Multimedia resource recommendation method, device, electronic device and storage medium

Abstract

The present disclosure relates to a multimedia resource recommendation method, device, electronic device and storage medium, and relates to the field of computer technology. In an embodiment of the present disclosure, the interests of a user account are integrated and analyzed through multimedia resources that the user account has interacted with, and the multimedia resources are analyzed based on the clustering of the multimedia resources determined according to similarity to identify which multimedia resources are the core interests of the user account. The similarity between the multimedia resources and the candidate multimedia resources can be used to identify which multimedia resources are similar to the candidate multimedia resources. Combining these two factors, it is naturally possible to analyze multimedia resources that are similar to the candidate multimedia resources and can represent the core interests of the user account from the multimedia resources that the user account has interacted with, so as to recommend multimedia resources to the user account, so that the recommended target multimedia resources are consistent with the core interests of the user account and meet the user's preferences, thereby improving the accuracy of the recommendation results.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to a multimedia resource recommendation method, device, electronic device, and storage medium. Background Art

[0002] With the rapid development of computer technology, intelligent recommendation plays an increasingly important role. The main task of intelligent recommendation is to obtain the user's preferences or interests in multimedia resources by analyzing the relevant information of the user account, and then recommend multimedia resources to the user account accordingly.

[0003] In related technologies, multimedia resource recommendation methods usually obtain multimedia resources that a user account has interacted with, and use convolutional neural networks (CNN) or recurrent neural networks (RNN) to predict the multimedia resource that is most likely to be interacted with next from multiple candidate multimedia resources, thereby recommending the multimedia resource to the user account.

[0004] If a user account has interacted with a large number of multimedia resources, the maximum pooling step in the CNN will ignore important portions of these multimedia resources and recurrent signals, failing to accurately capture the user account's core interests. However, RNNs tend to forget long-term features, so extracting a user account's interests from previously interacted multimedia resources encounters short-term bottlenecks, making it impossible to capture the user account's long-term interests, and any captured interests are also inaccurate. Consequently, the user features extracted by this method are inaccurate, leading to poor recommendation accuracy. Summary of the Invention

[0005] The present disclosure provides a multimedia resource recommendation method, device, electronic device, and storage medium, which can improve the accuracy of recommendation results. The technical solution of the present disclosure is as follows:

[0006] According to a first aspect of an embodiment of the present disclosure, a multimedia resource recommendation method is provided, comprising:

[0007] Acquire a multimedia resource set that the user account has interacted with, the multimedia resource set including a plurality of first multimedia resources;

[0008] Determining, from the multimedia resource set, a third multimedia resource corresponding to each candidate multimedia resource based on a first resource category parameter of each first multimedia resource, and similarities between the first multimedia resource and a second multimedia resource of the first multimedia resource and a plurality of candidate multimedia resources, wherein the first resource category parameter is used to indicate a possibility that the first multimedia resource can represent a corresponding category, the second multimedia resource is a multimedia resource in the multimedia resource set whose similarity to the first multimedia resource satisfies a first similarity condition, the similarity between the third multimedia resource and the candidate multimedia resource satisfies a second similarity condition, and the first resource category parameter of the third multimedia resource satisfies a parameter condition;

[0009] Based on the multiple third multimedia resources, multimedia resources are recommended to the user account.

[0010] In some embodiments, determining, from the multimedia resource set, a third multimedia resource corresponding to each candidate multimedia resource based on the first resource category parameter of each first multimedia resource, the first multimedia resource, and similarities between the second multimedia resource of the first multimedia resource and a plurality of candidate multimedia resources, includes:

[0011] For each candidate multimedia resource among the plurality of candidate multimedia resources, determining association information between the first multimedia resource and the second multimedia resource based on the first resource category parameter of each first multimedia resource, the first multimedia resource, and a similarity between the second multimedia resource of the first multimedia resource and the candidate multimedia resource;

[0012] According to the association information between the first multimedia resource and the second multimedia resource, a plurality of third multimedia resources corresponding to the candidate multimedia resource are determined from the multimedia resource set, and the association information corresponding to the plurality of third multimedia resources is not zero.

[0013] In some embodiments, determining association information between the first multimedia resource and the second multimedia resource based on the first resource category parameter of each first multimedia resource, the first multimedia resource, and the similarity between the second multimedia resource of the first multimedia resource and the candidate multimedia resource includes:

[0014] Acquire candidate association information between the first multimedia resource and the second multimedia resource according to the first resource category parameter of each first multimedia resource and the first similarity between the second multimedia resource of the first multimedia resource and the candidate multimedia resource;

[0015] determining, based on the candidate association information between the first multimedia resource and the second multimedia resource, a second resource category parameter of the first multimedia resource and a second similarity between the first multimedia resource and the candidate multimedia resource, wherein the second resource category parameter is a resource category parameter obtained by updating the first resource category parameter;

[0016] The association information between the first multimedia resource and the second multimedia resource is determined according to the second resource category parameter of the first multimedia resource, the second similarity, the first target value, and the candidate association information between the first multimedia resource and the second multimedia resource.

[0017] In some embodiments, determining the second resource category parameter of the first multimedia resource and the second similarity between the first multimedia resource and the candidate multimedia resource based on the candidate association information between the first multimedia resource and the second multimedia resource includes:

[0018] Based on the candidate association information between the first multimedia resource and the second multimedia resource, updating the representation information of the first multimedia resource to obtain target representation information of the first multimedia resource;

[0019] Based on the target representation information of the first multimedia resource, a first resource category parameter of the first multimedia resource is updated to obtain a second resource category parameter of the first multimedia resource;

[0020] Based on the target representation information of the first multimedia resource, a second similarity between the first multimedia resource and the candidate multimedia resource is determined.

[0021] In some embodiments, determining the association information between the first multimedia resource and the second multimedia resource based on the second resource category parameter of the first multimedia resource, the second similarity, the first target value, and the candidate association information between the first multimedia resource and the second multimedia resource includes:

[0022] determining importance indication information of the first multimedia resource according to the second resource category parameter of the first multimedia resource and the second similarity;

[0023] In response to the importance indication information of the first multimedia resource being greater than or equal to a first target value, using the candidate association information between the first multimedia resource and the second multimedia resource as the association information between the first multimedia resource and the second multimedia resource;

[0024] In response to the importance indication information of the first multimedia resource being less than a first target value, the association information between the first multimedia resource and the second multimedia resource is set to zero.

[0025] In some embodiments, the first target value is a value greater than the second target value in the importance indication information of each first multimedia resource, and the second target value is a minimum value of a target length and the number of the first multimedia resources.

[0026] In some embodiments, the method for obtaining the first resource category parameter of the first multimedia resource includes:

[0027] Determining a target category corresponding to the first multimedia resource;

[0028] Acquire average representation information of the first multimedia resources of the target category according to the representation information of each first multimedia resource belonging to the target category;

[0029] A third similarity between the representation information of the first multimedia resource and the average representation information is obtained, and the third similarity is used as a first resource category parameter of the first multimedia resource.

[0030] In some embodiments, the recommending multimedia resources to the user account based on the plurality of third multimedia resources includes:

[0031] Determining a target multimedia resource from the plurality of candidate multimedia resources based on the plurality of third multimedia resources and an interaction time sequence of the plurality of third multimedia resources;

[0032] The target multimedia resource is recommended to the user account.

[0033] In some embodiments, determining the target multimedia resource from the plurality of candidate multimedia resources based on the plurality of third multimedia resources and the interaction time sequence of the plurality of third multimedia resources includes:

[0034] Acquire an interaction sequence for each candidate multimedia resource, wherein the interaction sequence is composed of the plurality of third multimedia resources arranged in interaction time sequence;

[0035] Obtaining a matching degree between the interaction sequence of each candidate multimedia resource and each candidate multimedia resource;

[0036] The candidate multimedia resource with the greatest matching degree is used as the target multimedia resource.

[0037] In some embodiments, obtaining the matching degree between the interaction sequence of each candidate multimedia resource and each candidate multimedia resource includes:

[0038] For each candidate multimedia resource, obtaining first representation information corresponding to the plurality of third multimedia resources according to the plurality of third multimedia resources, where the first representation information is used to represent a set consisting of the plurality of third multimedia resources;

[0039] acquiring second representation information according to the importance indication information of each third multimedia resource and a change in the third multimedia resource in the interaction sequence, where the second representation information conforms to the change;

[0040] A matching degree between the interaction sequence of the candidate multimedia resource and the candidate multimedia resource is acquired according to the first representation information, the second representation information, and the representation information of the candidate multimedia resource.

[0041] In some embodiments, obtaining a matching degree between the interaction sequence of the candidate multimedia resource and the candidate multimedia resource based on the first representation information, the second representation information, and the representation information of the candidate multimedia resource includes:

[0042] Obtaining a Hadamard product of the first representation information and the representation information of the candidate multimedia resource;

[0043] Connecting the first representation information, the second representation information, the representation information of the candidate multimedia resource, and the Hadamard product to obtain third representation information;

[0044] A dimensionality reduction process is performed on the third representation information to obtain a matching degree between the interaction sequence of the candidate multimedia resource and the candidate multimedia resource.

[0045] According to a second aspect of an embodiment of the present disclosure, a multimedia resource recommendation device is provided, comprising:

[0046] An acquiring unit is configured to acquire a multimedia resource set that a user account has interacted with, wherein the multimedia resource set includes a plurality of first multimedia resources;

[0047] a determining unit configured to determine, from the multimedia resource set, a third multimedia resource corresponding to each candidate multimedia resource based on a first resource category parameter of each first multimedia resource, a similarity between the first multimedia resource and a second multimedia resource of the first multimedia resource and a plurality of candidate multimedia resources, wherein the first resource category parameter is used to indicate a possibility that the first multimedia resource can represent a corresponding category, the second multimedia resource is a multimedia resource in the multimedia resource set whose similarity to the first multimedia resource satisfies a first similarity condition, the similarity between the third multimedia resource and the candidate multimedia resource satisfies a second similarity condition, and the first resource category parameter of the third multimedia resource satisfies a parameter condition;

[0048] The recommendation unit is configured to perform multimedia resource recommendation to the user account based on the multiple third multimedia resources.

[0049] In some embodiments, the determining unit is configured to perform:

[0050] For each candidate multimedia resource among the plurality of candidate multimedia resources, determining association information between the first multimedia resource and the second multimedia resource based on the first resource category parameter of each first multimedia resource, the first multimedia resource, and a similarity between the second multimedia resource of the first multimedia resource and the candidate multimedia resource;

[0051] According to the association information between the first multimedia resource and the second multimedia resource, a plurality of third multimedia resources corresponding to the candidate multimedia resource are determined from the multimedia resource set, and the association information corresponding to the plurality of third multimedia resources is not zero.

[0052] In some embodiments, the determining unit is configured to perform:

[0053] Acquire candidate association information between the first multimedia resource and the second multimedia resource according to the first resource category parameter of each first multimedia resource and the first similarity between the second multimedia resource of the first multimedia resource and the candidate multimedia resource;

[0054] determining, based on the candidate association information between the first multimedia resource and the second multimedia resource, a second resource category parameter of the first multimedia resource and a second similarity between the first multimedia resource and the candidate multimedia resource, wherein the second resource category parameter is a resource category parameter obtained by updating the first resource category parameter;

[0055] The association information between the first multimedia resource and the second multimedia resource is determined according to the second resource category parameter of the first multimedia resource, the second similarity, the first target value, and the candidate association information between the first multimedia resource and the second multimedia resource.

[0056] In some embodiments, the determining unit is configured to perform:

[0057] Based on the candidate association information between the first multimedia resource and the second multimedia resource, updating the representation information of the first multimedia resource to obtain target representation information of the first multimedia resource;

[0058] Based on the target representation information of the first multimedia resource, a first resource category parameter of the first multimedia resource is updated to obtain a second resource category parameter of the first multimedia resource;

[0059] Based on the target representation information of the first multimedia resource, a second similarity between the first multimedia resource and the candidate multimedia resource is determined.

[0060] In some embodiments, the determining unit is configured to perform:

[0061] determining importance indication information of the first multimedia resource according to the second resource category parameter of the first multimedia resource and the second similarity;

[0062] In response to the importance indication information of the first multimedia resource being greater than or equal to a first target value, using the candidate association information between the first multimedia resource and the second multimedia resource as the association information between the first multimedia resource and the second multimedia resource;

[0063] In response to the importance indication information of the first multimedia resource being less than a first target value, the association information between the first multimedia resource and the second multimedia resource is set to zero.

[0064] In some embodiments, the first target value is the largest value in the importance indication information of each first multimedia resource, and the second target value is the minimum value of the target length and the number of first multimedia resources in the multimedia resource set.

[0065] In some embodiments, the method for obtaining the first resource category parameter of the first multimedia resource includes:

[0066] Determining a target category corresponding to the first multimedia resource;

[0067] Acquire average representation information of the first multimedia resources of the target category according to the representation information of each first multimedia resource belonging to the target category;

[0068] A third similarity between the representation information of the first multimedia resource and the average representation information is obtained, and the third similarity is used as a first resource category parameter of the first multimedia resource.

[0069] In some embodiments, the recommendation unit includes a determination module and a recommendation module;

[0070] The determining module is configured to determine a target multimedia resource from the plurality of candidate multimedia resources based on the plurality of third multimedia resources and an interaction time sequence of the plurality of third multimedia resources;

[0071] The recommendation module is configured to recommend the target multimedia resource to the user account.

[0072] In some embodiments, the determination module includes an acquisition submodule and a determination submodule;

[0073] The acquisition submodule is configured to execute an interaction sequence for acquiring each candidate multimedia resource, wherein the interaction sequence is composed of the plurality of third multimedia resources arranged in interaction time sequence;

[0074] The acquisition submodule is further configured to execute acquisition of a matching degree between the interaction sequence of each candidate multimedia resource and each candidate multimedia resource;

[0075] The determination submodule is configured to execute the process of selecting the candidate multimedia resource with the greatest matching degree as the target multimedia resource.

[0076] In some embodiments, the acquisition submodule is configured to perform:

[0077] For each candidate multimedia resource, obtaining first representation information corresponding to the plurality of third multimedia resources according to the plurality of third multimedia resources, where the first representation information is used to represent a set consisting of the plurality of third multimedia resources;

[0078] acquiring second representation information according to the importance indication information of each third multimedia resource and a change in the third multimedia resource in the interaction sequence, where the second representation information conforms to the change;

[0079] A matching degree between the interaction sequence of the candidate multimedia resource and the candidate multimedia resource is acquired according to the first representation information, the second representation information, and the representation information of the candidate multimedia resource.

[0080] In some embodiments, the acquisition submodule is configured to perform:

[0081] Obtaining a Hadamard product of the first representation information and the representation information of the candidate multimedia resource;

[0082] Connecting the first representation information, the second representation information, the representation information of the candidate multimedia resource, and the Hadamard product to obtain third representation information;

[0083] A dimensionality reduction process is performed on the third representation information to obtain a matching degree between the interaction sequence of the candidate multimedia resource and the candidate multimedia resource.

[0084] According to a third aspect of an embodiment of the present disclosure, an electronic device is provided, comprising: one or more processors; one or more memories for storing instructions executable by the one or more processors; wherein the processor is configured to execute the instructions to implement the above-mentioned multimedia resource recommendation method.

[0085] According to a fourth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided. When instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform the above-mentioned multimedia resource recommendation method.

[0086] According to a fifth aspect of an embodiment of the present disclosure, a computer program product is provided, comprising computer instructions, which implement the above-mentioned multimedia resource recommendation method when executed by a processor.

[0087] The technical solutions provided by the embodiments of the present disclosure bring at least the following beneficial effects:

[0088] In the embodiment of the present disclosure, the interests of the user account are integrated and analyzed through the multimedia resources that the user account has interacted with. The multimedia resources are clustered according to the similarity to analyze which multimedia resources are the core interests of the user account. The similarity between the multimedia resources and the candidate multimedia resources can be analyzed to determine which multimedia resources are similar to the candidate multimedia resources. Combining these two factors, it is naturally possible to analyze the multimedia resources that are similar to the candidate multimedia resources and can represent the core interests of the user account from the multimedia resources that the user account has interacted with. In this way, multimedia resources can be recommended to the user account, so that the recommended target multimedia resources are in line with the core interests of the user account and meet the user preferences, thereby improving the accuracy of the recommendation results.

[0089] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] The accompanying drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute an improper limitation of the present disclosure.

[0091] Figure 1 It is a structural diagram of a graph convolutional neural network according to an exemplary embodiment.

[0092] Figure 2 The figure is a schematic diagram showing an implementation environment of a multimedia resource recommendation method according to an exemplary embodiment.

[0093] Figure 3 The figure is a flowchart of a multimedia resource recommendation method according to an exemplary embodiment.

[0094] Figure 4 The figure is a flowchart of a multimedia resource recommendation method according to an exemplary embodiment.

[0095] Figure 5 The figure is a flowchart of a graph neural network according to an exemplary embodiment.

[0096] Figure 6 The figure is a flowchart of a multimedia resource recommendation method according to an exemplary embodiment.

[0097] Figure 7 The figure is a flowchart of a multimedia resource recommendation method according to an exemplary embodiment.

[0098] Figure 8 The figure is a flowchart of a multimedia resource recommendation method according to an exemplary embodiment.

[0099] Figure 9 is a schematic diagram of an experimental result according to an exemplary embodiment.

[0100] Figure 10 is a schematic diagram of an experimental result according to an exemplary embodiment.

[0101] Figure 11 is a schematic diagram of an experimental result according to an exemplary embodiment.

[0102] Figure 12 The figure is a block diagram of a multimedia resource recommendation device according to an exemplary embodiment.

[0103] Figure 13 The figure is a schematic structural diagram of an electronic device according to an exemplary embodiment.

[0104] Figure 14 The figure is a structural block diagram of a terminal according to an exemplary embodiment.

[0105] Figure 15 The figure is a schematic diagram showing the structure of a server according to an exemplary embodiment. DETAILED DESCRIPTION

[0106] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0107] It should be noted that the terms "first," "second," and the like in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the numbers used in this manner are interchangeable where appropriate so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of apparatus and methods consistent with certain aspects of the present disclosure as detailed in the appended claims.

[0108] The user account information involved in this disclosure may be information authorized by the user account or fully authorized by all parties.

[0109] The following are explanations of the terms involved in this disclosure.

[0110] (1) Figure

[0111] In the embodiments of the present disclosure, a graph refers to data stored in the form of a graph, which is also called graph data, graph model, graph representation, or graph structure data. A graph includes at least one node and at least one edge, each node having corresponding features, and edges are used to represent the connection relationship between different nodes. Optionally, a graph is defined as G = (V, E). Where G represents a graph, V represents the set of nodes in the graph, and E represents the set of edges in the graph. Optionally, the edges of the graph have weights, and the weights of the edges represent the properties of the connection relationship.

[0112] (2) Graph Neural Networks

[0113] Graph Neural Network (GNN) refers to a general term for models that apply neural networks to graphs. Graph Neural Networks include Graph Convolutional Networks (GCN), Graph Attention Networks, and the like. Graph Neural Networks are used to predict the category of a graph based on the structural features of the graph. Specifically, a graph neural network will include one or more feature extraction layers. The feature extraction layer is, for example, a Graph Convolution Layer (GCL). The feature extraction layer is used to extract the structural features of the graph. If two graphs are isomorphic, then the structural features obtained from the graphs of the two graphs will be similar after passing through the feature extraction layer. If the two graphs are heterogeneous, then the structural features obtained from the graphs of the two graphs will be different after passing through the feature extraction layer. Therefore, graph neural networks can map graph structures with isomorphic properties to the same representation domain and output the same category. For example, Figure 1 is a structural diagram of a graph convolutional neural network according to an exemplary embodiment. Figure 1 As shown in the figure, in a specific example, the input data of the graph convolutional neural network (GCN) is a graph, which includes nodes and edges, where the nodes are X and the edges are the adjacency matrix A. After each layer in the graph convolutional neural network, the data is processed by the Rectified Linear Unit (ReLU) function, and the last layer can obtain outputs.

[0114] (3) Graph Convolutional Network

[0115] A graph convolutional network (GCN) is a type of graph neural network that uses graph convolution. A GCN consists of at least one GCN layer. The GCN layer functions similarly to a feature extractor, extracting features from the graph and extracting structural features from the graph. Specifically, a GCN layer includes multiple convolution operators, also known as convolution kernels. A convolution kernel is essentially a weight matrix. The weights in this matrix are obtained during the model training phase. These weight matrices, formed from these trained weights, can be used to extract features from the input graph, enabling the GCN to make accurate predictions during application.

[0116] Optionally, the graph is input into the graph convolutional network in the form of matrix data, and the nodes, edges, and weights in the graph are represented by the values ​​in the matrix. For example, the features of the nodes in the graph are represented by an N×D-dimensional matrix X, where N represents the number of nodes in the graph, that is, the number of nodes included in the set V described in (1) above. D represents the dimension of the feature vector of each node. The eigenvalues ​​in the feature vector of the node are, for example, the values ​​of the attributes of the node. For example, if the node has three attributes, the feature vector of the node includes three attribute values ​​corresponding to the three attributes, and D is 3. The features of the edges in the graph are represented by an N×N-dimensional adjacency matrix A. For example, if there is an edge between two nodes, the weights of the corresponding edges of the two nodes in the adjacency matrix A are non-zero values. If there is no edge between the two nodes, the weights of the corresponding edges of the two nodes in the adjacency matrix A are 0. The adjacency matrix A can indicate the connection relationship between any two nodes in the N nodes. N is a positive integer.

[0117] (4) Graph convolution processing

[0118] The graph convolution layer performs feature extraction through graph convolution. Graph convolution is an operation that performs nonlinear transformations on input data. For the first convolutional layer of a graph convolutional network, the input data is the graph; for the second through last convolutional layers, the input data is the output of the previous convolutional layer.

[0119] Figure 2 FIG. 1 is a schematic diagram showing an implementation environment of a multimedia resource recommendation method according to an exemplary embodiment. Figure 2 The implementation environment includes a terminal 201, or the implementation environment includes a terminal 201 and a multimedia resource recommendation platform 202. The terminal 201 is connected to the multimedia resource recommendation platform 202 via a wireless network or a wired network.

[0120] Terminal 201 can be at least one of a smartphone, a game console, a desktop computer, a tablet computer, an e-book reader, an MP3 (Moving Picture Experts Group Audio Layer III) player or an MP4 (Moving Picture Experts Group Audio Layer IV) player, a laptop computer, an intelligent robot, or a self-service payment device. Terminal 201 has installed and running an application that supports multimedia resource recommendations. For example, the application can be a system application, an instant messaging application, a news push application, a shopping application, an online video application, or a social networking application.

[0121] For example, a user account can be logged into the terminal 201, and the user can operate on the terminal to interact with any multimedia resource. The terminal 201 can send the interaction record to the multimedia resource recommendation platform 202, and the multimedia resource recommendation platform 202 provides related services. The terminal 201 can also send the interaction record to other devices, and the other devices can provide related services. The multimedia resource recommendation platform 202 can retrieve the interaction record from the other devices.

[0122] For example, when the terminal 201 has a multimedia resource recommendation requirement, it can send a recommendation request to the multimedia resource recommendation platform 202 , and the multimedia resource recommendation platform 202 recommends multimedia resources to the user account logged in on the terminal 201 in response to the recommendation request.

[0123] For example, the multimedia resource recommendation platform 202 can proactively recommend multimedia resources to a user account based on the relevant information of the user account. The terminal 201 receives the recommended multimedia resources and can display the multimedia resources.

[0124] The multimedia resource recommendation platform 202 includes at least one of a single server, multiple servers, a cloud computing platform, and a virtualization center. The multimedia resource recommendation platform 202 is configured to provide backend services for applications that support multimedia resource recommendation. Optionally, the multimedia resource recommendation platform 202 performs primary processing, while the terminal 201 performs secondary processing. Alternatively, the multimedia resource recommendation platform 202 performs secondary processing, while the terminal 201 performs primary processing. Alternatively, the multimedia resource recommendation platform 202 or the terminal 201 can each independently perform processing. Alternatively, the multimedia resource recommendation platform 202 and the terminal 201 utilize a distributed computing architecture for collaborative computing.

[0125] Optionally, the multimedia resource recommendation platform 202 includes at least one server 2021 and a database 2022. The database 2022 is used to store data. In an embodiment of the present application, the database 2022 can store sample images or sample facial images to provide data services for at least one server 2021.

[0126] A server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. A terminal can be, but is not limited to, a smartphone, tablet computer, laptop computer, desktop computer, smart speaker, smartwatch, etc.

[0127] Those skilled in the art will appreciate that the number of the terminals 201 and servers 2021 can be greater or lesser. For example, there can be only one terminal 201 or server 2021, or there can be dozens, hundreds, or even more terminals 201 or servers 2021. The present embodiment does not limit the number or device type of the terminals or servers.

[0128] Figure 3 is a flowchart of a multimedia resource recommendation method according to an exemplary embodiment. Figure 3 As shown, the multimedia resource recommendation method can be executed by an electronic device, including the following steps.

[0129] In step S31 , a multimedia resource set that the user account has interacted with is obtained, where the multimedia resource set includes a plurality of first multimedia resources.

[0130] In step S32, based on the first resource category parameter of each first multimedia resource, the similarity between the first multimedia resource and the second multimedia resource of the first multimedia resource and multiple candidate multimedia resources, the third multimedia resource corresponding to each candidate multimedia resource is determined from the multimedia resource set, the first resource category parameter is used to indicate the possibility that the first multimedia resource can represent the corresponding category, the second multimedia resource is a multimedia resource in the multimedia resource set whose similarity with the first multimedia resource satisfies the first similarity condition, the similarity between the third multimedia resource and the candidate multimedia resource satisfies the second similarity condition, and the first resource category parameter of the third multimedia resource satisfies the parameter condition.

[0131] In step S33, multimedia resources are recommended to the user account based on the multiple third multimedia resources.

[0132] In the embodiment of the present disclosure, the interests of the user account are integrated and analyzed through the multimedia resources that the user account has interacted with. The multimedia resources are clustered according to the similarity to analyze which multimedia resources are the core interests of the user account. The similarity between the multimedia resources and the candidate multimedia resources can be analyzed to determine which multimedia resources are similar to the candidate multimedia resources. Combining these two factors, it is naturally possible to analyze the multimedia resources that are similar to the candidate multimedia resources and can represent the core interests of the user account from the multimedia resources that the user account has interacted with. In this way, multimedia resources can be recommended to the user account, so that the recommended target multimedia resources are in line with the core interests of the user account and meet the user preferences, thereby improving the accuracy of the recommendation results.

[0133] In some embodiments, determining a third multimedia resource corresponding to each candidate multimedia resource from the multimedia resource set based on the first resource category parameter of each first multimedia resource, the similarity between the first multimedia resource and the second multimedia resource of the first multimedia resource, and a plurality of candidate multimedia resources, includes:

[0134] For each candidate multimedia resource among the plurality of candidate multimedia resources, determining association information between the first multimedia resource and the second multimedia resource based on the first resource category parameter of each first multimedia resource, the first multimedia resource, and a similarity between the second multimedia resource of the first multimedia resource and the candidate multimedia resource;

[0135] According to the association information between the first multimedia resource and the second multimedia resource, a plurality of third multimedia resources corresponding to the candidate multimedia resource are determined from the multimedia resource set, and the association information corresponding to the plurality of third multimedia resources is not zero.

[0136] In some embodiments, determining association information between the first multimedia resource and the second multimedia resource based on the first resource category parameter of each first multimedia resource, the first multimedia resource, the similarity between the second multimedia resource of the first multimedia resource and the candidate multimedia resource includes:

[0137] Acquire candidate association information between the first multimedia resource and the second multimedia resource according to the first resource category parameter of each first multimedia resource and the first similarity between the second multimedia resource of the first multimedia resource and the candidate multimedia resource;

[0138] Determining, based on the candidate association information between the first multimedia resource and the second multimedia resource, a second resource category parameter of the first multimedia resource and a second similarity between the first multimedia resource and the candidate multimedia resource, wherein the second resource category parameter is a resource category parameter obtained by updating the first resource category parameter;

[0139] The association information between the first multimedia resource and the second multimedia resource is determined according to the second resource category parameter of the first multimedia resource, the second similarity, the first target value, and the candidate association information between the first multimedia resource and the second multimedia resource.

[0140] In some embodiments, determining the second resource category parameter of the first multimedia resource and the second similarity between the first multimedia resource and the candidate multimedia resource based on the candidate association information between the first multimedia resource and the second multimedia resource includes:

[0141] Based on the candidate association information between the first multimedia resource and the second multimedia resource, updating the representation information of the first multimedia resource to obtain target representation information of the first multimedia resource;

[0142] Based on the target representation information of the first multimedia resource, a first resource category parameter of the first multimedia resource is updated to obtain a second resource category parameter of the first multimedia resource;

[0143] Based on the target representation information of the first multimedia resource, a second similarity between the first multimedia resource and the candidate multimedia resource is determined.

[0144] In some embodiments, determining the association information between the first multimedia resource and the second multimedia resource based on the second resource category parameter of the first multimedia resource, the second similarity, the first target value, and the candidate association information between the first multimedia resource and the second multimedia resource includes:

[0145] determining importance indication information of the first multimedia resource according to the second resource category parameter of the first multimedia resource and the second similarity;

[0146] In response to the importance indication information of the first multimedia resource being greater than or equal to a first target value, using the candidate association information between the first multimedia resource and the second multimedia resource as the association information between the first multimedia resource and the second multimedia resource;

[0147] In response to the importance indication information of the first multimedia resource being less than a first target value, the association information between the first multimedia resource and the second multimedia resource is set to zero.

[0148] In some embodiments, the first target value is a value greater than the second target value in the importance indication information of each first multimedia resource, and the second target value is a minimum value of the target length and the number of the first multimedia resources.

[0149] In some embodiments, a method for obtaining the first resource category parameter of the first multimedia resource includes:

[0150] Determining a target category corresponding to the first multimedia resource;

[0151] Obtaining average representation information of the first multimedia resources of the target category according to the representation information of each first multimedia resource belonging to the target category;

[0152] A third similarity between the representation information of the first multimedia resource and the average representation information is obtained, and the third similarity is used as a first resource category parameter of the first multimedia resource.

[0153] In some embodiments, recommending multimedia resources to the user account based on the plurality of third multimedia resources includes:

[0154] Determining a target multimedia resource from the plurality of candidate multimedia resources based on the plurality of third multimedia resources and an interaction time sequence of the plurality of third multimedia resources;

[0155] The target multimedia resource is recommended to the user account.

[0156] In some embodiments, determining the target multimedia resource from the plurality of candidate multimedia resources based on the plurality of third multimedia resources and the interaction time sequence of the plurality of third multimedia resources includes:

[0157] Acquire an interaction sequence for each candidate multimedia resource, where the interaction sequence consists of the plurality of third multimedia resources arranged in interaction time sequence;

[0158] Obtaining a matching degree between the interaction sequence of each candidate multimedia resource and each candidate multimedia resource;

[0159] The candidate multimedia resource with the greatest matching degree is used as the target multimedia resource.

[0160] In some embodiments, obtaining the matching degree between the interaction sequence of each candidate multimedia resource and each candidate multimedia resource includes:

[0161] For each candidate multimedia resource, obtaining first representation information corresponding to the plurality of third multimedia resources based on the plurality of third multimedia resources, where the first representation information is used to represent a set consisting of the plurality of third multimedia resources;

[0162] acquiring second representation information according to the importance indication information of each third multimedia resource and a change in the third multimedia resource in the interaction sequence, where the second representation information conforms to the change;

[0163] A matching degree between the interaction sequence of the candidate multimedia resource and the candidate multimedia resource is obtained according to the first representation information, the second representation information, and the representation information of the candidate multimedia resource.

[0164] In some embodiments, obtaining a matching degree between the interaction sequence of the candidate multimedia resource and the candidate multimedia resource according to the first representation information, the second representation information, and the representation information of the candidate multimedia resource includes:

[0165] Obtaining a Hadamard product of the first representation information and the representation information of the candidate multimedia resource;

[0166] Connecting the first representation information, the second representation information, the representation information of the candidate multimedia resource, and the Hadamard product to obtain third representation information;

[0167] A dimensionality reduction process is performed on the third representation information to obtain a matching degree between the interaction sequence of the candidate multimedia resource and the candidate multimedia resource.

[0168] Figure 4 FIG1 is a flowchart of a multimedia resource recommendation method according to an exemplary embodiment. Figure 4 , the method comprising:

[0169] Step S41: The electronic device obtains a plurality of candidate multimedia resources and a multimedia resource set that the user account has interacted with, where the multimedia resource set includes a plurality of first multimedia resources.

[0170] The first multimedia resource and the candidate multimedia resources may be any one of text, audio, image, and video. Accordingly, the application scenario of the embodiment of the present disclosure may be any one of a text recommendation scenario, an audio recommendation scenario, an image recommendation scenario, and a video recommendation scenario. For example, a short video is a type of video, and the above-mentioned first multimedia resource and the candidate multimedia resources may be short videos. The application scenario of the embodiment of the present disclosure may be a short video recommendation scenario. For another example, the above-mentioned first multimedia resource and the candidate multimedia resources may be news, and the application scenario of the embodiment of the present disclosure may be a news recommendation scenario. For another example, the above-mentioned first multimedia resource and the candidate multimedia resources may be advertisements, and the application scenario of the embodiment of the present disclosure may be an advertisement recommendation scenario.

[0171] In an embodiment of the present disclosure, the electronic device can predict the multimedia resource that the user account is most likely to interact with next based on the first multimedia resource that the user account has interacted with, thereby recommending the multimedia resource to the user account. It can be understood that the first multimedia resource that the user account has interacted with can usually express the interests of the user account to a certain extent, and generally user accounts tend to interact with similar multimedia resources. For example, in a video recommendation scenario, suppose a user loves electronic games, and when watching multimedia resources, he prefers to watch live videos or short videos of electronic games. When the candidate multimedia resources include news videos, sports videos, and electronic game videos, the user account is more likely to watch electronic game videos next than to watch other types of videos.

[0172] A candidate multimedia resource is a multimedia resource to be recommended, and the multimedia resource that the user account is most likely to interact with next is one determined from the candidate multimedia resources. Accordingly, in the current application scenario, when multimedia resources need to be recommended for a user account, and multiple candidate multimedia resources are included, the electronic device obtains the first multimedia resource that the user account has interacted with, analyzes the user account's interests based on this, and then determines which candidate multimedia resource the user account is likely to interact with next based on the user account's interests.

[0173] A user account can interact with the first multimedia resource in a variety of ways, and the interaction methods can also vary in different application scenarios. For example, the interaction methods can include likes, clicks, forwarding, sharing, downloading, and comments. The specific method or methods of interaction used by the user account to interact with the first multimedia resource can be set by relevant technical personnel according to needs, and the present embodiment does not limit this.

[0174] Specifically, the electronic device can determine the likelihood of the user account interacting with each candidate multimedia resource to select the multimedia resource with which the user account is most likely to interact for recommendation. The multimedia resource with the highest likelihood of interaction is referred to below as the target multimedia resource. Thus, the method provided in the embodiments of the present disclosure is used to determine the target multimedia resource from the candidate multimedia resources and recommend it to the user account.

[0175] The electronic device can obtain the multimedia resource set that the user account has interacted with in a variety of ways. Several possible ways are provided below. The embodiments of the present disclosure can adopt any of the ways, which are not specifically limited here.

[0176] In method one, the electronic device can extract the set of multimedia resources that the user account has interacted with from the local recorded data. When the user account interacts with a first multimedia resource, the terminal where the user account is located can send the relevant data of this interaction to the electronic device. After the electronic device receives the relevant data of this interaction, it can record the first multimedia resource that the user account has interacted with. Of course, the electronic device can also record the interaction time of this interaction, and the interaction time can be recorded in the electronic device corresponding to the first multimedia resource. In this way, when it is necessary to recommend multimedia resources for the user account later, the set of multimedia resources that the user account has interacted with in the recorded data can be obtained.

[0177] In method 2, the electronic device can obtain from the database the set of multimedia resources that the user account has interacted with. When the user account interacts with a first multimedia resource, the terminal where the user account is located can send the relevant data of this interaction to the electronic device, and the electronic device stores it in the database. Alternatively, the user account can interact with a first multimedia resource through a target application, and the terminal where the user account is located can send the relevant data of this interaction to the server corresponding to the target application, and the server corresponding to the target application stores it in the database. Subsequently, the electronic device can obtain from the database the set of multimedia resources that the user account has interacted with.

[0178] The set of multimedia resources that the user account has interacted with may include all the first multimedia resources that the user account has interacted with, and may also include some of the first multimedia resources that the user account has interacted with. In some embodiments, the electronic device may obtain all the first multimedia resources that the user account has interacted with, and use this as data basis for subsequent processing. In other embodiments, the electronic device may obtain the first multimedia resources that the user account has interacted with within the target historical time period, and use this as data basis for subsequent processing. The first multimedia resources that the user account has interacted with within the target historical time period are the some of the first multimedia resources that the user account has interacted with.

[0179] In some embodiments, the first multimedia resource that the user account has interacted with can be stored corresponding to the identification information of the user account. When it is necessary to obtain the first multimedia resource that the user account has interacted with, the electronic device can obtain the first multimedia resource stored corresponding to the identification information based on the identification information of the user account. The identification information is used to uniquely identify the user account. The identification information can include multiple forms, and the embodiments of the present disclosure can adopt any type of identification information. For example, the identification information can be the account ID (Identification) of the user account, the account name of the user account, the name of the user account, etc. Of course, the identification information can also be other information, such as a certain identity credential of the user account or the terminal ID of the user account, etc. The embodiments of the present disclosure are not limited to this.

[0180] It should be noted that step S41 is described as an example in which the electronic device obtains the candidate multimedia resources. In some embodiments, the candidate multimedia resources can be obtained later. In other embodiments, the candidate multimedia resources may not be obtained, but may be resources that are already available when training is required. This is not limited in the present embodiment.

[0181] Step S42: For each candidate multimedia resource among the multiple candidate multimedia resources, the electronic device determines the association information between the first multimedia resource and the second multimedia resource based on the first resource category parameter of each first multimedia resource, the first multimedia resource and the similarity between the second multimedia resource of the first multimedia resource and the candidate multimedia resource.

[0182] The electronic device obtains a set of multimedia resources that a user account has interacted with. The set of multimedia resources that the user account has interacted with can reflect the interests of the user account. In the disclosed embodiments, when recommending multimedia resources, it is necessary to analyze the user account's relatively important interests from the set of multimedia resources that the user account has interacted with, and from these, analyze the interests related to the candidate multimedia resources to be recommended. The relatively important interests of the user account are referred to as core interests, which are used to distinguish them from marginal interests.

[0183] A user account may have different levels of interest in different multimedia resources. For example, a user account may be particularly interested in some multimedia resources, that is, the user account may have a deep or great interest in these multimedia resources. As time goes by, the user account's interest in these multimedia resources may not change much, so these multimedia resources can also reflect the user account's long-term interests. The user account may be slightly interested in other multimedia resources, that is, the user account's interest in these other multimedia resources is relatively shallow or small. The user account may only be interested in them for a period of time. Subsequently, as time goes by, the user account's interest in these other multimedia resources gradually fades, or even no longer has interest, so these other multimedia resources can also reflect the user account's short-term interests. Here, interests can be divided according to the depth or size of the interest. Specifically, deep or large interests are called core interests, and shallow or small interests are called marginal interests.

[0184] It is understandable that the user account may interact more frequently with multimedia resources related to core interests, while it may interact less frequently with multimedia resources of marginal interests. The following analysis of the clustering of first multimedia resources can be conducted through the similarity between the first multimedia resources in the multimedia resource set. It is understandable that similar first multimedia resources are more likely to cluster together, while first multimedia resources with large differences will be farther apart, thereby making it possible to explicitly and more easily distinguish the core interests of the user account from the marginal interests. Furthermore, the method provided in the embodiment of the present disclosure aims to extract the core interests of the user account, thereby recommending target multimedia resources that meet the core interests to the user account based on the core interests, so that the user account is more likely to interact with the target multimedia resources, thereby improving the recommendation conversion rate and success rate.

[0185] The following is an explanation of the nouns involved.

[0186] The first resource category parameter is used to indicate the possibility that the first multimedia resource can represent the corresponding category. That is, when clustering analysis is performed on the first multimedia resources, the distance between similar first multimedia resources is smaller, and the distance between dissimilar first multimedia resources is larger. In this way, first multimedia resources that are similar to each other will be clustered together to form a class, and the first multimedia resources in the multimedia resource set can also form multiple classes, each class being used to represent a category. For multiple first multimedia resources of a category, there will usually be a cluster center, which refers to the center of each cluster or class, and the sum of the distances of all first multimedia resources in the cluster or class to the cluster center is the smallest. Assuming that the sum of the distances of all first multimedia resources in a class to the cluster center is d, the sum of the distances of all first multimedia resources in the class to other locations is greater than d, and d is greater than zero.

[0187] It is understandable that the first multimedia resource that serves as the cluster center can better and more accurately represent the corresponding category. In the disclosed embodiment, the first resource category parameter can be used to determine which first multimedia resource is more representative of the corresponding category. The first multimedia resource that can represent the corresponding category is the first multimedia resource that is expected to be determined to represent the user's core interests.

[0188] The second multimedia resource is a multimedia resource in the multimedia resource set whose similarity to the first multimedia resource meets the first similarity condition. When analyzing any first multimedia resource in the multimedia resource set, we would like to analyze whether there are many first multimedia resources similar to it in the multimedia resource set, so that this first multimedia resource and similar first multimedia resources may be able to reflect the core interests of the user. For example, if a user is interested in a first multimedia resource, he may often interact with this first multimedia resource, and naturally the number of such first multimedia resources in the multimedia resource set that the user account has interacted with will be relatively large. Then, when analyzing a first multimedia resource, the second multimedia resource can be determined first based on the similarity between the first multimedia resources, and then the association information between the first multimedia resource and the second multimedia resource can be further determined by combining multiple factors.

[0189] In the embodiment of the present disclosure, when determining the association information between the first multimedia resource and the second multimedia resource, three factors can be comprehensively considered: the first resource category parameter of the first multimedia resource, the first similarity between the first multimedia resource and the candidate multimedia resource, and the second similarity between the second multimedia resource and the candidate multimedia resource. Through these three factors, we hope to be able to make the determined association information reflect whether the first multimedia resource is more important, so as to reflect the core interests of the user, and also reflect whether the corresponding category of the first multimedia resource is similar to the category of the candidate multimedia resource. The embodiment of the present disclosure aims to select a third multimedia resource from the multimedia resource set that can represent the core interests of the user and is relatively relevant to the candidate multimedia resource. The third multimedia resource is more representative in determining whether the user is interested in the candidate multimedia resource.

[0190] The following is an explanation of the process of determining the second multimedia resource.

[0191] In some embodiments, for each first multimedia resource in a multimedia resource set, the electronic device can obtain the similarity between each first multimedia resource and other first multimedia resources in the multimedia resource set, and then determine the first multimedia resource in the multimedia resource set whose similarity meets the first similarity condition as the second multimedia resource of the multimedia resource.

[0192] In some embodiments, the similarity between first multimedia resources is measured by the similarity between representation information of the first multimedia resources. For each first multimedia resource that a user account has interacted with, the electronic device may embed the first multimedia resource to obtain representation information of the first multimedia resource. In one specific embodiment, the representation information may be an embedding vector.

[0193] For any two first multimedia resources in the multimedia resource set, the electronic device obtains the similarity between the representation information of the two first multimedia resources and determines the second multimedia resource of each first multimedia resource based on the similarity.

[0194] The similarity acquisition process can be implemented in a variety of ways. The similarity acquisition process can be understood as a similarity measurement process, and the acquisition process can be implemented using any similarity measurement function. For example, the similarity acquisition process can be implemented using any of cosine similarity, radial basis function (RBF) kernel, and attention mechanism. Of course, the similarity between the nodes can also be obtained using other similarity measurement functions, such as Jaccard similarity coefficient, Euclidean distance, etc., which are not specifically limited in the embodiments of the present disclosure.

[0195] In a specific possible embodiment, the electronic device may use weighted cosine similarity as a metric function to determine the similarity between the two first multimedia resources, which may be specifically implemented by the following formula 1.

[0196] Characterization information of the first multimedia resource j. S ij is the similarity between the first multimedia resource i and the first multimedia resource j. is the weight vector, which is related to the input wheat sign information and have the same dimension. i and j are multimedia resource identifiers. In the above example, the representation information and The dimension of is d, then the The dimension of is also d, where d is a positive integer. In some embodiments, the representation information of the first multimedia resource may be an embedding vector.

[0197] In some embodiments, the weight vector may be updated during the training process. In this way, the weight values ​​of different dimensions in the weight vector learned may be different, and features emphasizing different dimensions of the above-mentioned representation information are learned.

[0198] In some embodiments, the clustering of the first multimedia resources can be observed from multiple perspectives, and the similarities between the first multimedia resources observed from different perspectives may be different. Therefore, the electronic device can expand the similarity measurement function into multiple perspective versions, obtain the similarity between the two first multimedia resources from multiple perspectives, and combine the similarities of multiple perspectives as the final similarity. Specifically, φ weight vectors can be set, each weight vector corresponding to a perspective. The φ weight vectors are obtained in a learning phase or a training phase. After learning, the weight values ​​of different dimensions in the φ weight vectors may be different, so that the φ weight vectors can focus on features of different dimensions.

[0199] Accordingly, the electronic device may obtain multiple similarities based on multiple different weight vectors and the representation information of the two first multimedia resources, and then obtain the similarity between the two first multimedia resources based on the multiple similarities.

[0200] For the process of obtaining the similarity between the two first multimedia resources based on the multiple similarities, the electronic device can implement it through a variety of comprehensive methods. For example, the multiple similarities can be weighted to obtain the similarity between the two first multimedia resources. Wherein, when the weighted weight is set to 1, the weighted process is a summing process. When the weighted weight is set to the same value, and the sum of the multiple identical values ​​is 1, the weighted process is an averaging process. For another example, multiple similarities can be connected to obtain the similarity between the two first multimedia resources. The embodiments of the present disclosure do not limit which specific method is used to integrate multiple similarities.

[0201] In the above example of the cosine similarity function, the electronic device can calculate φ independent similarity matrices using the following formulas 2 and 3. Each similarity matrix is ​​based on the similarity of each perspective, and their average is used as the final similarity. Here, only the average of multiple similarities is used as the final similarity as an example. The electronic device can also concatenate the similarities of multiple perspectives to obtain the final similarity, or sum the similarities of multiple perspectives to obtain the final similarity. Of course, the final similarity can also be determined by other methods, and the embodiments of the present disclosure are not limited to this.

[0202]

[0203]

[0204] in, is the representation information of the two first multimedia resources under the δth perspective and The cosine similarity between them. cos() is the cosine function. ⊙ is the Hadamard product. is the representation information of the first multimedia resource i. is the representation information of the first multimedia resource j. ij is the similarity between the first multimedia resource i and the first multimedia resource j. is the weight vector for the δth perspective. ∑ is the summation function. Each perspective captures a portion of the semantics of the embedded vector and focuses on a different aspect. Measuring the connections between the first multimedia resources from multiple perspectives can improve expressiveness and obtain more accurate similarity.

[0205] The electronic device determines the similarity between the first multimedia resources and can determine whether one of the first multimedia resources is a second multimedia resource of the other first multimedia resource based on the similarity. It is understood that the higher the similarity between the first multimedia resources, the more likely the two first multimedia resources are to be classified into the same category, and the greater the likelihood that the two first multimedia resources are each other's second multimedia resources. The lower the similarity between the first multimedia resources, the more likely the two first multimedia resources are to be classified into different categories, and the less likely the two first multimedia resources are to be each other's second multimedia resources.

[0206] In some embodiments, a similarity threshold may be provided, and the similarity between the two first multimedia resources may be divided by the similarity threshold to determine whether the two first multimedia resources are each other's second multimedia resources. Specifically, the electronic device determines that the two first multimedia resources are each other's second multimedia resources in response to the similarity between the two first multimedia resources being greater than or equal to the similarity threshold. The electronic device determines that the two first multimedia resources are not each other's second multimedia resources in response to the similarity between the two first multimedia resources being less than the similarity threshold.

[0207] The similarity threshold can be set by relevant technical personnel according to needs, or it can be determined based on sparsity, and the embodiments of the present disclosure are not limited to this. For example, the similarity threshold can be set to 0.7. For another example, the relevant technical personnel set the sparsity, and the electronic device can determine the similarity threshold based on the sparsity and the similarity between each first multimedia resource. The strategy in which the similarity threshold is determined based on sparsity can be called a relative ranking strategy. Through sparsity, the numerical value ranked in a certain position in the similarity between each first multimedia resource can be used as the similarity threshold.

[0208] In some embodiments, the electronic device may obtain the square of the number of first multimedia resources in the multimedia resource set, obtain the product of the square and the sparsity, and use the similarity between the first multimedia resources with the largest product as the similarity threshold. Assuming that the number of first multimedia resources in the multimedia resource set is n and the sparsity is ε, the product is εn 2 Among the similarities between the first multimedia resources, the similarity with the largest product can be recorded as is the εnth similarity in S 2 Large value.

[0209] In some embodiments, the similarity threshold is determined based on the sparsity, and the sparsity is used to perform the sparsification. iiThe value range of is [-1, 1], and it can be normalized to [0, 1], and the similarity is converted into a positive number. Of course, normalization can also be performed, and the embodiments of the present disclosure are not limited to this. In a specific possible embodiment, simply normalizing it may make the clustering of the first multimedia resources more uniform, and the number of first multimedia resources of different categories may not differ much. By setting the sparsity for sparseness, the amount of calculation can be reduced, and the associations between some unimportant first multimedia resources can be filtered out, thereby enabling the subsequent information aggregation steps to focus on the most important part of the multimedia resource set. The two first multimedia resources are called multimedia resource pairs. The above method can consider the most important multimedia resource pairs and filter out unimportant multimedia resource pairs to determine the second first multimedia resource of each first multimedia resource from S. S refers to the above S ij The matrix composed of .

[0210] In some embodiments, the sparsity-based determination of the similarity threshold can be implemented using a relative ranking strategy. This makes the hyperparameters insensitive and does not disrupt the sparse distribution of the multimedia resource set. Specifically, the association indication information A between the first multimedia resources can be further obtained using Formula 4.

[0211]

[0212] Among them, A ij is the association indication information between the first multimedia resource i and the first multimedia resource j, is the εnth similarity in S 2 The largest value refers to the similarity threshold. n is the number of the first multimedia resources in the multimedia resource set, and ε is the sparsity, which is used to control the overall sparsity. Otherwise refers to the value that does not meet the Other cases, here refers to In the fourth formula, by setting the association indication information with similarity less than the similarity threshold to zero, the elements in S with similarity less than the similarity threshold are shielded. The similarity threshold is selected by ranking the metric values ​​(i.e., similarity) in S. ij , A ij If A is 1, the first multimedia resource i and the first multimedia resource j are each other's second multimedia resources. ij If it is 0, the first multimedia resource i and the first multimedia resource j are not each other's second multimedia resources.

[0213] By clustering the first multimedia resources that each user account has interacted with, it is easier to distinguish between their core interests and peripheral interests. Understandably, the cluster corresponding to the first multimedia resources representing core interests will contain more first multimedia resources with similar interests. The number of second multimedia resources associated with this first multimedia resource may be relatively large.

[0214] In some embodiments, the electronic device may convert the first multimedia resources that the user account has interacted with into a historical interaction sequence according to the interaction time sequence, where each element in the historical interaction sequence is the first multimedia resource performed at a specific time point. The first multimedia resource is further processed based on the historical interaction sequence.

[0215] In step S42, the electronic device performs cluster analysis on the first multimedia resources that the user account has interacted with based on metric learning. The first multimedia resources that the user account has interacted with can be understood as one-dimensional data. Arranging the first multimedia resources that the user account has interacted with in chronological order of interaction can form a one-dimensional historical interaction sequence. Some user accounts may have interacted with a large number of first multimedia resources, while other user accounts may have interacted with a small number of first multimedia resources. Therefore, the historical interaction sequence may be very long or relatively short.

[0216] The above description has clarified the process of determining the second multimedia resource. The following describes how to determine the association information between the first multimedia resource and the second multimedia resource.

[0217] The electronic device preliminarily determines a cluster distribution of the first multimedia resources based on the similarities between the first multimedia resources. The electronic device can then update the cluster distribution based on the first resource category parameters of the first multimedia resources and the similarities between the candidate multimedia resources and each of the first multimedia resources, thereby further highlighting first multimedia resources that are related to the candidate multimedia resources and that represent the core interests of the user account. This update process can be understood as an information aggregation process for perceiving the core interests of the user account related to the candidate multimedia resources.

[0218] In an embodiment of the present disclosure, the association information between the first multimedia resource and the second multimedia resource can be obtained using an attention mechanism. The attention mechanism analyzes which first multimedia resource or the relationship between two first multimedia resources requires more attention, thereby obtaining the association information between the first multimedia resource and the second multimedia resource. This association information can be understood as an attention score, which is used to indicate the importance of the relationship between the first multimedia resource and the second multimedia resource.

[0219] Specifically, step S42 can be implemented through the following steps 1 to 3.

[0220] Step 1: The electronic device obtains candidate association information between the first multimedia resource and the second multimedia resource based on the first resource category parameter of each first multimedia resource and the first similarity between the second multimedia resource of the first multimedia resource and the candidate multimedia resource.

[0221] Specifically, when determining candidate association information, the possibility that the first multimedia resource can represent the corresponding category can be evaluated. If the first multimedia resource can represent the corresponding category, then the first multimedia resource can represent the core interest of the user account, and naturally it is more important in the subsequent prediction process. More attention needs to be paid to the first multimedia resource and the second multimedia resource closely related to it. The candidate association information of the first multimedia resource and the second multimedia resource can be relatively large. Considering that the embodiment of the present disclosure is intended to analyze the possibility of interaction between the user account and the candidate multimedia resource, when analyzing whether the second multimedia resource is important to the first multimedia resource, it is possible to focus on whether the second multimedia resource is related to the candidate multimedia resource. If so, the second multimedia resource is more important, and more information of the second multimedia resource needs to be transferred to the first multimedia resource so that the representation information of the first multimedia resource is more relevant to the candidate multimedia resource, and vice versa. Therefore, when determining candidate association information, the first resource category parameter of the first multimedia resource and the first similarity between the second multimedia resource and the candidate multimedia resource can be determined first, and then the two can be combined to obtain candidate association information of the first multimedia resource and the second multimedia resource.

[0222] It should be noted that, herein, the similarity between the second multimedia resource and the candidate multimedia resource is referred to as the first similarity, and the similarity between the first multimedia resource and the candidate multimedia resource is referred to as the second similarity.

[0223] For the first resource category parameter of the first multimedia resource, the target category corresponding to the first multimedia resource can be determined first, and then the average representation information of the first multimedia resources of the target category can be obtained based on the representation information of each first multimedia resource belonging to the target category. The third similarity between the representation information of the first multimedia resource and the average representation information is obtained, and the third similarity is used as the first resource category parameter of the first multimedia resource.

[0224] By performing cluster analysis on the first multimedia resource, multiple first multimedia resources similar to the first multimedia resource are considered to be in the same category. By comparing the representation information of the first multimedia resource with the average representation information of first multimedia resources belonging to the same category, the representativeness of the first multimedia resource is measured. By comparing the representation information of the first multimedia resource with the average representation information of the corresponding category, the likelihood that the first multimedia resource represents the corresponding category is quantified and accurately represented using the first resource category parameters. This provides a data basis for analyzing the importance of the first multimedia resource, thereby enabling more accurate analysis of the core interests of user accounts.

[0225] The likelihood that the first multimedia resource represents the corresponding category can be understood as an attention score of the first multimedia resource. In some embodiments, the attention score can be calculated by an attention network, which can be a neural network. The attention score can be calculated using the following formula 5.

[0226]

[0227] Among them, α i is the attention score of the first multimedia resource i, that is, the first resource category parameter of the first multimedia resource i, W c is the transformation matrix, || is the connection operator, and ⊙ represents the Hadamard product. c It is an attention network. is the representation information of the first multimedia resource i, and the representation information may be an embedded vector. is the average representation information of the category (target category) corresponding to the first multimedia resource i. is the average representation information of the corresponding category (target category).

[0228] In a specific possible embodiment, the neural network may include two fully connected layers and an activation layer. The activation layer may use any activation function. For example, in a specific example, the attention network ATTE c It is a two-layer feedforward neural network with LeakyReLU as the activation function.

[0229] Regarding the first similarity between the second multimedia resource and the candidate multimedia resource, by analyzing the first similarity between the second multimedia resource and the candidate multimedia resource, the candidate multimedia resource can be used to dynamically bring the set of multimedia resources that the user account has interacted with closer to the candidate multimedia resource, and whether each first multimedia resource is related to the candidate multimedia resource can be analyzed in a targeted manner, so as to more accurately analyze whether the candidate multimedia resource is related to the core interests of the user.

[0230] In some embodiments, the first similarity can be measured by the similarity between the representation information. Specifically, the electronic device can consider the representation information of the second multimedia resource. and representation information of candidate multimedia resources The correlation between the two is also called the first similarity. If the correlation between the second multimedia resource and the candidate multimedia resource is higher, then the information transmission of the second multimedia resource to the first multimedia resource during the information aggregation process will be more obvious, and vice versa. This is also to take into account that the first multimedia resource related to the candidate multimedia resource can play a role in the final prediction process, retaining relevant information while discarding irrelevant information, so that the prediction result is not affected by irrelevant information and is more accurate.

[0231] The first resource category parameter of the first multimedia resource can be understood as the attention score of the first multimedia resource. Similarly, the first similarity between the second multimedia resource and the candidate multimedia resource can be understood as the attention score of the second multimedia resource. The attention score can be calculated using an attention network, which can be a neural network. The attention score can be calculated using the following formula 6.

[0232]

[0233] Among them, β j is the attention score of the second multimedia resource j, W q is the transformation matrix, || is the connection operator, and ⊙ represents the Hadamard product. q It is an attention network. is the representation information of the second multimedia resource j, and the representation information may be an embedding vector. It is the representation information of the candidate multimedia resources.

[0234] Similarly, the attention network ATTE q It can include two fully connected layers and an activation layer. The activation layer can use any activation function. For example, in a specific example, the attention network ATTE q It is a two-layer feedforward neural network with LeakyReLU as the activation function.

[0235] When determining candidate association information between a first multimedia resource and a second multimedia resource, two factors can be comprehensively considered: the first resource category parameter of the first multimedia resource and the first similarity between the second multimedia resource and the candidate multimedia resource. For example, the first resource category parameter and the first similarity can be weighted. This weighting method can include summing or averaging. Of course, other methods can also be used to combine the two factors, such as multiplication, which is not limited in the present embodiment.

[0236] In one possible implementation, the electronic device can sum the first resource category parameter and the first similarity of the first multimedia resource to obtain candidate association information between the first multimedia resource and the second multimedia resource. The first resource category parameter of the first multimedia resource can be understood as the attention score of the first multimedia resource. The first similarity between the second multimedia resource and the candidate multimedia resource can be understood as the attention score of the second multimedia resource. Since it considers the relevance with the candidate multimedia resource, the candidate multimedia resource can be a multimedia resource to be recommended or a multimedia resource to be queried. Therefore, the attention score of the second multimedia resource can also be called the query score of the second multimedia resource. The two scores are added together, which can be used as the candidate association information between the first multimedia resource i and the second multimedia resource j.

[0237] In one specific embodiment, a first multimedia resource may correspond to multiple second multimedia resources. Different second multimedia resources may have different, and even significantly different, first similarities. When considering the two factors, the electronic device may also normalize the second multimedia resources corresponding to a first multimedia resource to make the contributions of different second multimedia resources comparable.

[0238] Specifically, all selections of the second multimedia resource j may be normalized using the softmax function according to the following formula 7.

[0239]

[0240] Among them, e ij Refers to the candidate association information between the first multimedia resource i and the second multimedia resource j, which can be understood as the attention score of the relationship between the first multimedia resource i and the second multimedia resource j. Softmax is a normalization function. When j==i, the self-loop weight is obtained. The neighborhood of the first multimedia resource i Including the first multimedia resource i itself, is the neighborhood of the first multimedia resource i, including the first multimedia resource i and the second multimedia resource j similar to the first multimedia resource i. Under the condition of self-loop propagation, α i Controls how much information the first multimedia resource i can receive, and β j Control how much information the second multimedia resource j can send.

[0241] Step 2: The electronic device determines a second resource category parameter of the first multimedia resource and a second similarity between the first multimedia resource and the candidate multimedia resource based on the candidate association information between the first multimedia resource and the second multimedia resource, where the second resource category parameter is a resource category parameter after updating the first resource category parameter.

[0242] The electronic device determines the candidate association information between the first multimedia resource and the second multimedia resource, which can represent the information transfer situation from the second multimedia resource to the first multimedia resource. According to the candidate association information, the representation information of the first multimedia resource can also be updated, so that the representation information of the first multimedia resource includes the information transfer situation, so that the representation information of the first multimedia resource that is more related to the candidate multimedia resource can be more reflected, and the representation information of similar first multimedia resources can also be made more similar.

[0243] After the representation information of the first multimedia resource is updated, the association information between the first multimedia resource and the second multimedia resource can naturally be updated. In this way, the association information between the first multimedia resource and the second multimedia resource and the representation information of the first multimedia resource are both updated. The updated representation information is referred to as target representation information. The electronic device can determine the first multimedia resource that can represent the core interest of the user. The first multimedia resource that can represent the core interest is referred to as the third multimedia resource, and the target representation information of the third multimedia resource is obtained. The specific process can be found in step 2 and step 3 below.

[0244] The electronic device can update the representation information of the first multimedia resource based on the candidate association information between the first multimedia resource and the second multimedia resource to obtain the target representation information of the first multimedia resource, and then update the first resource category parameter of the first multimedia resource based on the target representation information of the first multimedia resource to obtain the second resource category parameter of the first multimedia resource, and then determine the second similarity between the first multimedia resource and the candidate multimedia resource based on the target representation information of the first multimedia resource.

[0245] When the second resource category parameter and the second similarity are determined based on the candidate association information, the candidate association information of the first multimedia resource and the second multimedia resource is determined. Naturally, based on the candidate association information, the representation information of the first multimedia resource can be updated so that the representation information of the first multimedia resource can be more clearly consistent with the candidate association information, and more accurate representation information can be obtained. If the representation information of the first multimedia resource changes, the parameters related to the first multimedia resource will naturally change accordingly. By updating the relevant parameters of the first multimedia resource through the determined candidate association information, and then performing subsequent calculation steps based on the updated parameters, it can be ensured that the subsequent calculation process can fully consider the distribution of the multimedia resource set and the relationship between each first multimedia resource, and obtain more accurate results.

[0246] In the process of acquiring the representation information of the first multimedia resource, it is assumed that the multimedia resource set is Where n is the number of first multimedia resources in the multimedia resource set (i.e., the length of the historical interaction sequence of the user account), and d is the dimension of the representation information of each first multimedia resource. The representation information of the first multimedia resource is the original representation information. The updated target representation information is represented as The dimension of the target representation information is d′, where the dimension of the target representation information may be the same as or different from the dimension of the representation information.

[0247] The candidate association information e of the first multimedia resource and the second multimedia resource ij The importance of the first multimedia resource i on its second multimedia resource j can be mapped, and the normalized candidate association information e ij It is used to calculate the linear combination with the representation information of the first multimedia resource to obtain the final representation information of each first multimedia resource, that is, the target representation information.

[0248] In a specific possible embodiment, the target representation information of the first multimedia resource can be calculated using the following formula 8.

[0249]

[0250] in, is the target representation information of the first multimedia resource i, is the representation information of the second multimedia resource j. aggregate is an aggregation function. W is a weight vector. σ is a nonlinear function. It is the neighborhood of the first multimedia resource i, that is, the set composed of the second multimedia resources j of the first multimedia resource i.

[0251] Among them, the aggregation function can be any one of the averaging function, the sum function, the maximum value, and the GRU function. Of course, the aggregation function can also use other functions, and the embodiments of the present disclosure are not limited to this.

[0252] In some embodiments, the target representation information can also be obtained by processing the representation information from multiple different perspectives. For example, the aggregation function is a summation function, and φ independent perspectives are used to obtain the information. φ independent attention mechanisms perform the transformation above and then concatenate their representation information to produce the following output embedding representation:

[0253]

[0254] Among them, || means concatenation, is the normalized attention coefficient calculated by the δth attention mechanism, W δ is the weight matrix of the corresponding input linear transformation. In the summation method, the output of each first multimedia resource The dimension of is φd′. If the average is taken, the dimension is d′.

[0255] Step 3: The electronic device determines the association information between the first multimedia resource and the second multimedia resource based on the second resource category parameter of the first multimedia resource, the second similarity, the first target value, and the candidate association information between the first multimedia resource and the second multimedia resource.

[0256] In the above information aggregation process, the fusion of implicit interest signals to explicit interest signals is completed. This can be understood as a pooling process, which aims to reasonably narrow the scope of data to be processed. Electronic devices can use pooling methods to further extract fused information. The electronic device can extract the first multimedia resource (that is, the third multimedia resource) representing the core interest from the multimedia resource set based on the importance of each first multimedia resource, and perform subsequent prediction processes through the third multimedia resource.

[0257] The electronic device aggregates information through the connection between the first multimedia resources and the second similarity between the first multimedia resources and the candidate multimedia resources, so that the second resource category parameters of the first multimedia resources and the second similarity between the first multimedia resources and the candidate multimedia resources are more accurate after the information aggregation, which can be helpful for analyzing which first multimedia resources are related to and more important than the candidate multimedia resources. Further, through information aggregation, subsequent predictions can be made more accurate, thereby improving the success rate and conversion rate of recommendations.

[0258] When the electronic device determines the possibility that each first multimedia resource is the third multimedia resource, the possibility may be determined according to the importance of the first multimedia resource, and the importance of the first multimedia resource may be represented by importance indication information of the first multimedia resource.

[0259] In some embodiments, the electronic device may determine the importance indication information of the first multimedia resource based on the second resource category parameter of the first multimedia resource and the second similarity. Then, the association information between the first multimedia resource and the second multimedia resource is determined based on the size relationship between the importance indication information and the target value. Specifically, the electronic device may, in response to the importance indication information of the first multimedia resource being greater than or equal to the first target value, use the candidate association information between the first multimedia resource and the second multimedia resource as the association information between the first multimedia resource and the second multimedia resource. The electronic device may, in response to the importance indication information of the first multimedia resource being less than the first target value, set the association information between the first multimedia resource and the second multimedia resource to zero.

[0260] By updating the association information between the first multimedia resource and the second multimedia resource, the updated data can better reflect the connection between the first multimedia resources. Based on this, it is possible to analyze whether each first multimedia resource is more important and whether it can play an important role in analyzing whether the candidate multimedia resource meets the core interests of the user account. Based on this, the association information between the first multimedia resource and the second multimedia resource is determined, and the relationship between unimportant first multimedia resources in the multimedia resource set is filtered out, and only the most important relationship is retained. In this way, a third multimedia resource that can better reflect the core interests of the user account can be found, thereby ensuring the accuracy of subsequent recommendation results.

[0261] In some embodiments, the importance indication information of the first multimedia resource can also be implemented using an attention mechanism. After obtaining the importance indication information of the first multimedia resource, the association information between the first multimedia resource and the second multimedia resource can be determined to evaluate the possibility that the first multimedia resource is the third multimedia resource.

[0262] In a specific possible embodiment, the electronic device can determine the possibility that the first multimedia resource is the third multimedia resource by using the following three formulas: i , which is the importance indication information of the first multimedia resource.

[0263]

[0264]

[0265]

[0266] in, is the attention score of the first multimedia resource i, that is, the second resource category parameter of the first multimedia resource i. Refers to the second similarity between the first multimedia resource i and the candidate multimedia resource. γ i is the importance indication information of the first multimedia resource i. is the target representation information of the first multimedia resource i. c and W q is a weight vector, which can be the same as or different from the weight vector in the process of obtaining the first resource category parameter and the first similarity. ⊙ represents the Hadamard product. ATTE c and ATTE q It is an attention network. It is the average representation information calculated based on the target representation information of each first multimedia resource belonging to the category corresponding to the first multimedia resource i (that is, the target category).

[0267] In the attention mechanism, except for the input being aggregated cluster information, its parameters are consistent with those in the interest fusion layer. That is, Formulas 10 through 12 are similar to Formulas 5 through 8 above, except that Formulas 5 through 8 above are used to determine the relevant parameters of the first multimedia resource i and the second multimedia resource j to determine the association information between the first multimedia resource i and the second multimedia resource j. Formulas 10 through 12, however, are used to determine the importance indication information for the first multimedia resource i. Therefore, the parameters of the second multimedia resource j are replaced by the parameters of the first multimedia resource i.

[0268] After determining the possibility that the first multimedia resource is the third multimedia resource γ i Then, the multimedia resource set can be sampled according to the probability, and a target number of first multimedia resources representing the corresponding category can be used as the third multimedia resource. The first multimedia resource representing the corresponding category is the first multimedia resource on the cluster center of the category.

[0269] The process of determining the third multimedia resource can be understood as a pooling process, and the target number can be the pooling length, or the minimum value between the pooling length and the number of first multimedia resources in the multimedia resource set.

[0270] The first target value is the value in the importance indication information of each first multimedia resource that is greater than the second target value, and the second target value is the minimum of the target length and the number of first multimedia resources in the multimedia resource set. By using a relative ranking strategy, a few representative first multimedia resources that represent the user's core interests are selected, rather than using all first multimedia resources for prediction. This can reduce computational complexity and produce prediction results that are more consistent with the user's core interests.

[0271] Assuming a given pooling length m<n, select the m clusters with the highest scores and include them in the pooled class. In order to determine the class distribution after pooling, a class distribution matrix E is necessary. Since the attention score e in the interest fusion layer ij Considering the affiliation between the first multimedia resource and the corresponding category, therefore, under the assumption that ij The matrix composed of is the class distribution matrix E, and then considering the above two points at the same time, the class distribution matrix after sampling is recorded as It can be obtained by the following formula 12.

[0272]

[0273] After extracting the important interests, the associated indication information A determined by the above formula 4 and the class distribution matrix after sampling can be used to Process according to the following formula 13 to obtain new associated indication information To ensure the connectivity strength between every two clusters.

[0274]

[0275] In Formula 13, if there is any common first multimedia resource in the cluster distribution corresponding to the original association indication information or the first multimedia resources in any two clusters are neighbors, then any two clusters c(i) and c(j) after pooling are connected. Therefore, the connection strength between clusters is determined by the membership relationship of the first multimedia resources. and edge weight A is determined.

[0276] In this way, through the above steps, a compact and coarsened distribution representing a stronger interest signal of the user account is obtained, and all the first multimedia resources in the compact and coarsened distribution are third multimedia resources.

[0277] Step S43: The electronic device determines, from the multimedia resource set, a plurality of third multimedia resources corresponding to the candidate multimedia resource based on the association information between the first multimedia resource and the second multimedia resource, where the association information corresponding to the plurality of third multimedia resources is not zero.

[0278] Step S42 and step S43 are the processes of determining the third multimedia resource corresponding to each candidate multimedia resource from the multimedia resource set based on the first resource category parameter of each first multimedia resource, the similarity between the first multimedia resource and the second multimedia resource of the first multimedia resource and multiple candidate multimedia resources, wherein the similarity between the third multimedia resource and the candidate multimedia resource satisfies the second similarity condition, and the first resource category parameter of the third multimedia resource satisfies the parameter condition.

[0279] The determination of the third multimedia resource takes into account three factors: the first resource category parameter of the first multimedia resource, the second similarity between the first multimedia resource and the candidate multimedia resource, and the first similarity between the second multimedia resource and the candidate multimedia resource. The above only uses the example of first determining the association information between the first multimedia resource and the second multimedia resource through these three factors, and then determining the third multimedia resource based on the association information between the first multimedia resource and the second multimedia resource. By determining the association information between the first multimedia resource and the second multimedia resource through these three factors, the unimportant connections between the first multimedia resources can be filtered out, and the important connections between the first multimedia resources can be retained. In this way, the connections between the more important first multimedia resources are retained, and the third multimedia resource can be accurately screened out, so that more accurate recommendation results can be obtained for multimedia resource recommendation.

[0280] In some embodiments, the electronic device may not determine the association information between the first multimedia resource and the second multimedia resource, but may directly obtain the importance indication information of each first multimedia resource and select the third multimedia resource based on the relationship between the importance indication information and the importance threshold. Of course, other methods may also be used, and the present disclosure is not limited to this.

[0281] The association information between the first multimedia resources has been determined in the above manner. The association information between the first multimedia resources representing the core interest is non-zero, while the other ones are 0. The electronic device can select the first multimedia resource representing the core interest (that is, the third multimedia resource) based on the association information.

[0282] Step S44: The electronic device determines a target multimedia resource from the multiple candidate multimedia resources according to the multiple third multimedia resources and the interaction time sequence of the multiple third multimedia resources.

[0283] After the electronic device obtains the third multimedia resource representing the core interest, it can use the third multimedia resource as a basis for predicting recommended multimedia resources to perform a prediction step, which is a step of determining a target multimedia resource based on the third multimedia resource.

[0284] Considering that the interests of the user account may change over time, the changes in the interests of the user account over time can be analyzed by combining the first multimedia resource that has been interacted with and the interaction time.

[0285] Under the combined influence of the external environment and internal cognition, the core interests of a user account may continue to evolve. A user account may develop an interest in various sports over a period of time, and a need for books at another time. In order to provide more relevant historical information for the final representation of interest, the temporal order relationship between interests can also be considered. In a manner in which the first multimedia resources that have been interacted with are converted into a historical interaction sequence, the interaction time sequence of each first multimedia resource is also the relative position in the historical behavior sequence, and the interaction time sequence of the third multimedia resource is also the relative position sequence in the historical behavior sequence. The third multimedia resources after clustering and information fusion are flattened into a sequence. Each element in the sequence represents an enhanced interest signal.

[0286] The process of predicting the multimedia resources that the user account is most likely to interact with next based on the first multimedia resource that the user account has interacted with in the order of interaction time, and thus recommending the multimedia resources to the user account can be called a sequential recommendation process. This sequential recommendation aims to use the historical behavior sequence of the user account to predict the next behavior of the user account. This sequential recommendation process can be applied to any recommendation scenario, such as news recommendation, video recommendation, or advertising recommendation. Unlike traditional recommendation tasks that model the preferences of user accounts in a static manner, the sequential recommendation process can capture the dynamic preferences of user accounts over time, that is, it can obtain changes in the interests of user accounts over time. For example, a user account may be interested in various sports for a period of time, and like to read books at another time.

[0287] Specifically, the electronic device can obtain the interaction sequence of each candidate multimedia resource, and the interaction sequence is composed of the multiple third multimedia resources arranged in the order of interaction time. Then the electronic device obtains the matching degree between the interaction sequence of each candidate multimedia resource and each candidate multimedia resource, and uses the candidate multimedia resource with the greatest matching degree as the target multimedia resource. Taking the time sequence of the user account's interaction with the first multimedia resource into consideration, the interaction sequence can be used to indicate both the core interests of the user account and the changes in the core interests of the user account. It can quickly and accurately determine whether the candidate multimedia resource is consistent with the core interests of the user account and the changes in the core interests, and obtain accurate recommendation results.

[0288] The matching degree acquisition process can be achieved through the following steps 1 to 3.

[0289] Step 1: For each candidate multimedia resource, the electronic device obtains first representation information corresponding to the multiple third multimedia resources based on the multiple third multimedia resources. The first representation information is used to represent a set consisting of the multiple third multimedia resources.

[0290] The core interests of the user are aggregated to obtain the core interest expression of the user, which is also the first representation information. By characterizing the set as a whole, an expression that can represent the core interests of the user can be obtained. Specifically, for the set composed of the third multimedia resources, the interest characteristics of the user account can be obtained by weighting the representation information of the third multimedia resources. The interest characteristics can be represented at the set level. That is, the first characterization information, which can be obtained by the following formula 14.

[0291]

[0292] The weight is the pooled score (importance indication information) of each first multimedia resource γ i . Readout is a readout function. The readout function can be a function such as Mean, Sum, Max, etc. In a specific example, a simple sum function can be used to ensure permutation invariance. Of course, other functions can also ensure permutation invariance, and the embodiment of the present disclosure does not limit which specific function is used. This set-level representation is provided to the final prediction layer to better learn the importance weight of each retained first multimedia resource.

[0293] Step 2: The electronic device obtains second representation information according to the importance indication information of each third multimedia resource and the change of the third multimedia resource in the interaction sequence, and the second representation information conforms to the change.

[0294] Any known sequential recommendation method can be used to model the interaction sequence composed of third multimedia resources. Assuming a single sequential recommendation model is used to model the evolution of interests, the evolution method can be as shown in Formula 15: m third multimedia resources are used to evolve the core interests of the user account.

[0295]

[0296] in, It is the evolution result of the sequential recommendation model on the interaction sequence, which is also the second representation information. is target representation information of m third multimedia resources.

[0297] In some embodiments, the above-mentioned second representation information can be implemented through an interest evolution model. The electronic device can input the importance indication information of each third multimedia resource and the interaction sequence into the interest evolution model and output the second representation information. The AUGRU is the interest evolution model.

[0298] Step 3: The electronic device obtains a matching degree between the interaction sequence of the candidate multimedia resource and the candidate multimedia resource according to the first representation information, the second representation information, and the representation information of the candidate multimedia resource.

[0299] After obtaining the core interest expression, the representation information of the next interactive multimedia resource is also evolved. Through the above two representation information and the representation information of the candidate multimedia resources, it is possible to determine whether the candidate multimedia resource meets the user's core interest and the evolution of the user's core interest.

[0300] Through the first representation information, the second representation information that conforms to the changes in core interests, and the representation information of the candidate multimedia resources, it is possible to comprehensively and accurately analyze whether the candidate multimedia resources conform to the user's core interests and the changes in core interests, providing a data basis for multimedia resource recommendations.

[0301] The set-level representation of the interest extraction layer and the evolution output of the interest evolution layer are used as the user's current interests and are connected with the representation information of the candidate multimedia resources. Given a cascaded dense representation vector (referring to the data obtained by cascading the representation information of the third multimedia resource), a fully connected layer can be used to automatically learn the combination of representation information. A two-layer feedforward neural network is used as the prediction function Pred to estimate the probability of the user's next interaction with the candidate multimedia resources, and the multimedia resource with the greatest probability is determined as the target multimedia resource.

[0302] In some embodiments, the electronic device can obtain the Hadamard product of the first representation information and the representation information of the candidate multimedia resource, and then connect the first representation information, the second representation information, the representation information of the candidate multimedia resource and the Hadamard product to obtain third representation information, and then perform dimensionality reduction processing on the third representation information to obtain the matching degree between the interaction sequence of the candidate multimedia resource and the candidate multimedia resource.

[0303] By obtaining the Hadamard product, the first representation information and the representation information of the candidate multimedia resources can be completely retained, and the connection between the third multimedia resource of the current core interest and the candidate multimedia resources can be reflected, and then the third representation information is obtained by splicing it with the first representation information, the second representation information and the representation information of the candidate multimedia resources. In this way, the third representation information includes the set of third multimedia resources, the candidate multimedia resources, the connection between the set and the candidate multimedia resources, and the changes in the user's core interests. Based on this, dimensionality reduction processing is performed, and this information is comprehensively considered to obtain a matching degree that comprehensively considers multiple angles. The matching degree is more accurate, and multimedia resource recommendations are made in this way, which can improve the accuracy of the recommendation results.

[0304] The dimensionality reduction process can be implemented in various ways. In some embodiments, the dimensionality reduction process can be implemented by a multimedia resource prediction model, where the electronic device inputs the third representation information into the multimedia resource prediction model and outputs a matching degree between the interaction sequence of the candidate multimedia resource and the candidate multimedia resource.

[0305] For example, the dimensionality reduction process can be achieved through a fully connected layer. Assuming that the multimedia resource prediction model can be a two-layer feedforward neural network, the result of the multimedia resource prediction model can be as follows: Figure 5 As shown, the two hidden layers can be implemented using fully connected layers. The fully connected layer multiplies the input data by the weights, and then sums the product with the bias to obtain the output. This operation process is similar to the convolution operation performed by the 1x1 convolution layer. Therefore, the above-mentioned multimedia resource prediction model can also be implemented using convolution layers, which is not limited in the present embodiment.

[0306] The above process of determining the target multimedia resource can be implemented by the following formula 16.

[0307]

[0308] Here, Pred refers to a prediction function, which can be expressed as a sequential recommendation model, which is also a multimedia resource prediction model.

[0309] The network output layer of the multimedia resource prediction model can be a softmax layer, which can convert the above matching degree into a probability, making the results of multiple candidate multimedia resources more comparable. The softmax layer can be preceded by an L2 layer, which is used to regularize the results of the previous layer to obtain the matching degree.

[0310] It should be noted that the determination of the third multimedia resource, the determination of the second representation information, and the acquisition of the matching degree can all be achieved through a network or model. Through the training process of the network or model, the network or model can have accurate data processing capabilities, and then the trained network or model can be used to quickly and accurately process relevant data to improve data processing efficiency and accuracy, thereby improving recommendation efficiency and accuracy.

[0311] Step S45: The electronic device recommends the target multimedia resource to the user account.

[0312] After the electronic device determines the target multimedia resource, it can send the target multimedia resource to the target account to recommend the target multimedia resource to the target account. By processing the set of multimedia resources that the user account has interacted with, the user's core interests are extracted, and the recommendation process refers to the changes in core interests over time. Therefore, the recommendation results can be more in line with the needs of the user account, the number of interactions of the user account can be increased, the conversion rate of the recommended target multimedia resources can be increased, the recommendation success rate can be increased, and the recommendation effect can be better.

[0313] Step S44 and step S45 are the process of recommending multimedia resources to the user account based on the multiple third multimedia resources. This process takes into account the fact that the user's interests may change over time. The interaction time sequence of the third multimedia resources is combined in the prediction process, so that the determined target multimedia resources also conform to the user's interest changes, and are more in line with the current interests of the user account, and the recommendation results are more accurate.

[0314] In the embodiment of the present disclosure, the interests of the user account are integrated and analyzed through the multimedia resources that the user account has interacted with. The multimedia resources are clustered according to the similarity to analyze which multimedia resources are the core interests of the user account. The similarity between the multimedia resources and the candidate multimedia resources can be analyzed to determine which multimedia resources are similar to the candidate multimedia resources. Combining these two factors, it is naturally possible to analyze the multimedia resources that are similar to the candidate multimedia resources and can represent the core interests of the user account from the multimedia resources that the user account has interacted with. In this way, multimedia resources can be recommended to the user account, so that the recommended target multimedia resources are in line with the core interests of the user account and meet the user preferences, thereby improving the accuracy of the recommendation results.

[0315] In some embodiments, the electronic device can convert the first multimedia resource into a graph, and then integrate and analyze the interests of the user account in the form of a graph. Specifically, after the electronic device obtains the first multimedia resource that the user account has interacted with, it can generate a first graph based on the first multimedia resource that the user account has interacted with. The multimedia resource set will be mapped as a node in the graph, and the interests of the user account will be determined based on the first graph. The generation process can be understood as the process of converting the first multimedia resource that the user account has interacted with into the first graph, that is, the graph construction process. In this conversion process, the first multimedia resource is converted into a node in the first graph, and the similarity between the first multimedia resources is converted into the connection between the nodes, so as to analyze the connection between the first multimedia resources in the form of a graph. Subsequently, the edges of the graph are determined for the connection between the first multimedia resources, which can explicitly and more easily distinguish the core interests and marginal interests of the user account, and explicitly integrate and analyze the interests of the user account at different times.

[0316] Below through Figure 6 The illustrated embodiment illustrates a process of implementing multimedia resource recommendation by converting the multimedia resource into a graph and processing the graph. Figure 6 FIG1 is a flowchart of a multimedia resource recommendation method according to an exemplary embodiment. Figure 6 , the method comprising:

[0317] Step S61: The electronic device obtains a plurality of candidate multimedia resources and a multimedia resource set that the user account has interacted with, where the multimedia resource set includes a plurality of first multimedia resources.

[0318] This step S61 is similar to the above step S61 and will not be described in detail here.

[0319] Step S62: The electronic device generates a first graph based on the multimedia resource set. The first graph includes nodes for representing the first multimedia resources and edges for representing similarities between two first multimedia resources.

[0320] During analysis, the electronic device generates a first graph based on the first multimedia resources that the user account has interacted with, and subsequently determines the user account's interests based on the first graph. This generation process can be understood as the process of converting the first multimedia resources that the user account has interacted with into the first graph, that is, the graph construction process. During this conversion process, the first multimedia resources are converted into nodes in the first graph, and the similarities between the first multimedia resources are converted into connections between the nodes, thereby analyzing the connections between the nodes in the form of a graph, which can explicitly integrate and analyze the user account's interests at different times.

[0321] The process of generating the first graph is a graph construction process. The constructed first graph includes nodes and edges. The first graph can be understood as an initial graph. The representation information of the nodes and edges in the first graph will be processed or updated subsequently, so that the updated graph can more clearly express the core interests of the user account.

[0322] The following steps 1 and 2 are used to illustrate the process of generating nodes and edges of the first graph.

[0323] Step 1: The electronic device generates nodes in a first graph based on the first multimedia resources that the user account has interacted with, where each node is used to represent a first multimedia resource.

[0324] A node can also be referred to as a vertex, and each vertex is a first multimedia resource that the user account has interacted with. In step 1, the electronic device can convert all first multimedia resources that the user account has interacted with into nodes, thereby obtaining nodes in the first graph. The number of nodes in the first graph is the same as the number of first multimedia resources in the multimedia resource set. In other words, after the electronic device obtains the first multimedia resources that the user account has interacted with, it uses these first multimedia resources to analyze the interests of the user account.

[0325] In some embodiments, for each first multimedia resource that a user account has interacted with, the electronic device may embed the first multimedia resource to obtain representation information of the first multimedia resource, and use the representation information of the first multimedia resource as the representation information of the node representing the first multimedia resource in the first graph. In one specific possible embodiment, the representation information may be an embedding vector.

[0326] In a specific possible embodiment, the first graph can be understood as an initial graph, and the first graph can be an undirected graph. The undirected graph can be represented as in, It is node data or node set, which can also be called vertex data or vertex set. Each node (or vertex) in is denoted as v, and is used to represent a first multimedia resource that the user account has interacted with. and Where n is the number of nodes in the first graph, which is a positive integer. This number of nodes is the same as the first multimedia resource in the multimedia resource set. || is a modulo operation. Yes Performing modulo operation, we get For each node v, a representation information can be associated with each node, which is recorded as R dThe R in R refers to a real number, which means that the elements of each dimension in the representation information are real numbers. d The d in is the dimension of the information representation, which is a positive integer. ε and A are edge data. Where ε is the set of edges, A∈R n×n Is an adjacency matrix, each element A in the adjacency matrix i，j Used to indicate the slave node v i Points to node v j The weight of the edge. In an undirected graph, A is a symmetric matrix, A i，j With A j，i Equal. In a directed graph, the two can be different. Each edge can be recorded as (i, j, A i，j ), indicating whether node i is related to node j. (i, j, A i，j )∈ε.

[0327] In the embodiment of the present disclosure, the weight of the edge in the initial first graph is an initial value, and the weight of the edge needs to be updated in the subsequent processing process. The process of determining the weight of the edge is actually the process of determining the adjacency matrix A, and also the process of determining the edge set.

[0328] Step 2: The electronic device generates an edge in the first graph based on the similarity between two nodes in the first graph.

[0329] After the nodes in the first graph are determined, the electronic device can analyze whether any two nodes are similar. If two nodes are similar, there can be an edge between them, and the two nodes are neighbors. The distance between the two nodes in the first graph is relatively close. Subsequently, based on the distribution of nodes in the first graph, nodes that are clustered together are likely similar. If the number of clustered nodes is large, then these nodes are likely to be core interest nodes of the user account.

[0330] It should be noted that the process of generating the edges in the first graph in step 2 is the same as the above Figure 4 The process of determining the second multimedia resource in step S42 of the illustrated embodiment is similar. The second multimedia resource is a multimedia resource in the multimedia resource set whose similarity to the first multimedia resource satisfies the first similarity condition. In generating the first graph, this is correspondingly mapped to whether there are edges between nodes. Second multimedia resources that meet the first similarity condition can be considered neighbor nodes of nodes in the first graph.

[0331] In this step two, the electronic device can initialize the edges in the first graph through the nodes in the first graph. It can be understood that if the two first multimedia resources are relatively similar, the two nodes used to represent the two first multimedia resources are also relatively similar, and the representation information of the nodes should also be relatively similar. In some embodiments, when analyzing the similarity between nodes, the electronic device can measure the similarity between the two nodes by determining the similarity between the representation information of the two nodes. In this step two, the electronic device can generate the edges in the first graph through the similarity between the two nodes, so that the edges in the first graph can reflect the similarity between the two first multimedia resources.

[0332] This first graph will be further processed to aggregate information across the nodes in the graph and analyze the core interests of the user account. This means that the representations of the nodes in the first graph will be updated, and the graph structure will evolve as the representations are updated. This means that the connections between nodes will change as the representations are updated.

[0333] Specifically, the second step can be achieved through the following steps 2.1 and 2.2.

[0334] Step 2.1: For every two nodes in the first graph, the electronic device obtains the similarity between the representation information of the two nodes.

[0335] The similarity acquisition process can be implemented in a variety of ways. The similarity acquisition process can be understood as a similarity measurement process, and the acquisition process can be implemented using any similarity measurement function. For example, the similarity acquisition process can be implemented using any of cosine similarity, radial basis function (RBF) kernel, and attention mechanism. Of course, the similarity between the nodes can also be obtained using other similarity measurement functions, such as Jaccard similarity coefficient, Euclidean distance, etc., which are not specifically limited in the embodiments of the present disclosure.

[0336] In a specific possible embodiment, the electronic device may use weighted cosine similarity as a metric function to determine the similarity between the two nodes, which may be implemented by the above formula 1. Accordingly, the definitions of the parameters in the above formula 1 are: is the representation information of node i. is the representation information of node j. S ij is the similarity between node i and node j. i and j are node identifiers.

[0337] In some embodiments, the weight vector can be updated during training. That is, during training, the weight vector is updated as the graph generated based on the sample multimedia resource is processed. The weight values ​​of different dimensions in the learned weight vector may differ, thereby learning features that emphasize different dimensions of the aforementioned representation information.

[0338] In some embodiments, the first graph can be observed from multiple perspectives, and the similarities between nodes observed from different perspectives may be different. Therefore, the electronic device can expand the similarity measurement function into multiple perspective versions, obtain the similarity between two nodes from multiple perspectives, and combine the similarities of multiple perspectives as the final similarity. Specifically, φ weight vectors can be set, each weight vector corresponding to a perspective. The φ weight vectors are obtained in a learning phase or a training phase. After learning, the weight values ​​of different dimensions in the φ weight vectors may be different, so that the φ weight vectors can focus on features of different dimensions. That is, the same as the way of observing the similarity between the first multimedia resources from multiple perspectives in the above step S42.

[0339] Accordingly, in step 2.1, the electronic device may obtain multiple similarities based on multiple different weight vectors and the representation information of the two nodes, and then obtain the similarity between the two nodes based on the multiple similarities.

[0340] For the process of obtaining the similarity between the two nodes based on the multiple similarities, the electronic device can implement it through a variety of comprehensive methods. For example, the multiple similarities can be weighted to obtain the similarity between the two nodes. When the weighted weight is set to 1, the weighting process is a summing process. When the weighted weight is set to the same value, and the sum of the multiple same values ​​is 1, the weighting process is an averaging process. For another example, multiple similarities can be connected to obtain the similarity between the two nodes. The embodiments of the present disclosure do not limit the specific method used to integrate multiple similarities.

[0341] In the above example of the cosine similarity function, the electronic device can calculate φ independent similarity matrices using the above formulas 2 and 3, each similarity matrix being based on the similarity of each perspective, and taking their average value as the final similarity. Here, only the example of taking the average value of multiple similarities as the final similarity is used for illustration. The electronic device can also connect the similarities of the multiple perspectives to obtain the final similarity, or sum the similarities of the multiple perspectives to obtain the final similarity, etc. Of course, the final similarity can also be determined in other ways, which is not limited in the present embodiment. Accordingly, some parameters in the above formulas 2 and 3 are defined as: is the representation information of the two nodes at the δth perspective and The cosine similarity between . is the representation information of node i. is the representation information of node j. S ij is the similarity between node i and node j. Each perspective captures a portion of the semantics from the embedded vector and focuses on different aspects. Measuring the connections between nodes in the graph from multiple perspectives can improve expressiveness and obtain more accurate similarity.

[0342] Step 2.2: The electronic device determines an edge between the two nodes based on the similarity.

[0343] The electronic device determines the similarity between the nodes and can determine whether there is an edge between the nodes and the weight of the edge based on the similarity. It can be understood that if the similarity between the nodes is high, there may be an edge between the two nodes. If the similarity between the nodes is low, there may not be an edge between the two nodes.

[0344] The weight of the edge between two nodes can be set to a non-zero value to indicate that there is an edge between the two nodes, that is, the element corresponding to the edge between the two nodes in the adjacency matrix is ​​a non-zero value. The weight of the edge between two nodes can be set to 0 to indicate that there is no edge between the two nodes, that is, the element corresponding to the edge between the two nodes in the adjacency matrix is ​​zero.

[0345] It should be noted that the edge between the two nodes determined in step 2.2 is the initial state of the edge in the first graph, which is used to determine the neighbor nodes of the node in subsequent calculations. In subsequent calculations, the initial weight of the edge will be updated again.

[0346] In some embodiments, a similarity threshold may be set to divide the similarity between two nodes and determine the initial weight of the edge between the two nodes. To distinguish the edge weights at the beginning of the first graph construction from the new weights calculated subsequently, the edge weights determined in step 2.2 are referred to as initial weights.

[0347] Specifically, in response to the similarity between the two nodes being greater than or equal to a similarity threshold, the electronic device sets the initial weight of the edge between the two nodes to a first value, where the first value is a positive number. In response to the similarity between the two nodes being less than the similarity threshold, the electronic device sets the weight of the edge between the two nodes to zero.

[0348] In a possible implementation provided by an embodiment of the present disclosure, the first value can be set to 1, and the similarity between the nodes is not used to distinguish which nodes with edges are closer. This part of the content can be further determined through subsequent processing steps. That is, in response to the similarity between the two nodes being greater than or equal to the similarity threshold, the electronic device sets the initial weight of the edge between the two nodes to 1. In response to the similarity between the two nodes being less than the similarity threshold, the electronic device sets the initial weight of the edge between the two nodes to 0.

[0349] It should be noted that, in the above step S42, whether the two first multimedia resources are each other's second multimedia resources is determined by the relationship between the similarity between the two first multimedia resources and the similarity threshold. In the implementation in the form of a graph, the above first multimedia resources are mapped to nodes, and the second multimedia resources are also mapped to neighboring nodes of the nodes. The process of determining whether they are each other's second multimedia resources based on the above similarity is also mapped to the process of determining whether the initial weight of the edge between the two nodes is zero.

[0350] The above-mentioned similarity threshold can be set by relevant technical personnel according to needs, or it can be determined based on sparsity, and the embodiments of the present disclosure are not limited to this. For example, the similarity threshold can be set to 0.7. For another example, the relevant technical personnel set the sparsity, and the electronic device can determine the similarity threshold based on the sparsity and the similarity between the nodes in the first graph. The strategy in which the similarity threshold is determined based on sparsity can be called a relative ranking strategy. Through sparsity, the numerical value ranked in a certain noun in the similarity between the nodes in the first graph can be used as the similarity threshold.

[0351] In some embodiments, the electronic device may obtain the product of the square of the number of nodes in the first graph and the sparsity, and use the similarity between the nodes in the first graph that has the largest product as the similarity threshold. Assuming that the number of nodes in the first graph is n and the sparsity is ε, the product is εn 2 The similarity between the nodes in the first graph with the largest product can be recorded as is the εnth similarity in S 2 Large value.

[0352] In some embodiments, the graph is sparsified by the sparsity ε in order to determine the similarity threshold based on the sparsity. Considering that the adjacency matrix elements should be non-negative, the cosine similarity S calculated according to the metric ijThe value range of is [-1, 1], so it needs to be normalized. Simply normalizing it will not impose any constraints on the sparsity of the first graph, and will produce a fully connected adjacency matrix. By setting this sparsity to perform graph sparsification, the amount of calculation can be reduced, and some unimportant edges can be filtered out, so that the subsequent information aggregation steps can focus on the most important part of the first graph. The two nodes are called node pairs. The above method can consider the most important connected node pairs and filter out unimportant node pairs to extract the symmetric sparse non-negative adjacency matrix A from S. S refers to the above S ij The matrix composed of .

[0353] In some embodiments, the similarity threshold is determined based on sparsity and can be implemented through a relative ranking strategy. This makes the hyperparameters insensitive and does not destroy the sparse distribution of the graph. Specifically, the symmetric sparse non-negative adjacency matrix A can be obtained by the above formula four. The symmetric sparse non-negative adjacency matrix A also corresponds to the association indication information A between the first multimedia resources in the above step S42. Accordingly, the definition of some parameters in formula four can be: n is the number of nodes, ε is the sparsity, which is used to control the overall sparsity of the generated graph. In formula four, by setting the weights of the edges whose similarity is less than the similarity threshold to zero, those elements in S that are less than the similarity threshold are shielded, and the selection of the similarity threshold is obtained by ranking the metric values ​​(that is, the similarity) in S.

[0354] This strategy differs from the absolute threshold strategy for the entire graph and the relative ranking strategy for node neighborhoods. Setting an absolute threshold to remove smaller elements in the adjacency matrix can prevent the generation of a graph or the generation of a complete graph. This can occur because improper hyperparameter settings can cause the distribution of metric values ​​to change as the representation information is continuously updated, ultimately preventing the generation of a graph. This strategy also allows for node degrees to vary, enabling subsequent information aggregation to extract the core interests of user accounts.

[0355] By representing the first multimedia resources that each user account has interacted with as a graph, it is easier to distinguish between its core interests and marginal interests. For a node in the graph, if the first multimedia resource represented by the node can represent the core interests of the user account, the node can be called a core interest node. If the first multimedia resource represented by the node represents the marginal interests of the user account, the node can be called a marginal interest node. It can be understood that a core interest node is a node with more similar interests in the graph. The core interest node usually has a higher degree, and the higher the frequency of occurrence of the similar interests of the core interest node, the larger the subgraph can be formed based on the core interest node. In this way, through the graph construction step, an a priori framework is constructed, that is, neighbor nodes are more similar, that is, there is an edge between two nodes, and the two nodes are neighbor nodes to each other, then the two nodes are relatively similar. The embodiment of the present disclosure aims to determine a dense subgraph formed by core interest nodes, and the dense subgraph can reflect the core interests of the user account.

[0356] In some embodiments, the electronic device can convert the first multimedia resources that the user account has interacted with into a historical interaction sequence in chronological order, where each element in the historical interaction sequence is a first multimedia resource performed at a specific point in time. Accordingly, each node in the first graph corresponds to a first multimedia resource in the historical interaction sequence.

[0357] In step S62, the electronic device converts the first multimedia resources that the user account has interacted with into a first graph based on metric learning. The first multimedia resources that the user account has interacted with can be understood as one-dimensional data. Arranging the first multimedia resources that the user account has interacted with in chronological order of interaction can form a one-dimensional historical interaction sequence. Some user accounts may have interacted with a large number of first multimedia resources, while other user accounts may have interacted with a small number of first multimedia resources. Therefore, the historical interaction sequence may be very long or relatively short. Therefore, user accounts with different numbers of first multimedia resources in the multimedia resource set can all convert the first multimedia resources they have interacted with into a first graph, and analyze the interests of the user account based on the first graph. In other words, the electronic device can convert the first multimedia resources that the user account has interacted with into a historical interaction sequence in chronological order of interaction. Based on the historical interaction sequence, the electronic device generates nodes in the first graph. Then, the electronic device determines edges in the first graph based on the similarity between two nodes in the first graph.

[0358] Step S63: For each candidate multimedia resource among the multiple candidate multimedia resources, determine the weight of the first edge based on the first node parameter of each node in the first graph, the similarity between the node and its neighbor nodes and the candidate multimedia resource, where the first edge is the edge that the neighbor node points to the node, and the first node parameter is used to indicate the possibility that the node is the centroid of the node cluster.

[0359] After the electronic device generates the first graph, it can update the first graph based on the specific conditions of each edge in the first graph and the similarities between the candidate multimedia resources and the nodes in the first graph, so that the updated first graph can more clearly indicate the center of mass position of the node cluster in the graph, that is, the location of the core interest node of the user account. This update process can be understood as an information aggregation process, which is used to perceive the core interests of the user account related to the candidate multimedia resources (i.e., the node at the center of the node cluster).

[0360] The first resource category parameter of each first multimedia resource can be the first node parameter of each node in the first graph. The first resource category parameter is used to indicate the likelihood that the first multimedia resource represents the corresponding category. The set of first multimedia resources of the same category is mapped to a node cluster, and the first node parameter is used to indicate the likelihood that the node is the centroid of the node cluster.

[0361] The center of mass refers to the center of mass, which is an imaginary point on a material system where the mass is believed to be concentrated. It generally refers to the average position of the mass distribution. The center of mass of a node cluster refers to the average position of the distribution of all nodes in the node cluster. If a node is the center of mass of a node cluster, the sum of the distances from all nodes in the node cluster to the node is less than the sum of the distances from all nodes in the node cluster to any other node in the node cluster. That is, the sum of the distances from the nodes in the node cluster to the center of mass is the smallest. In this way, if a node is the center of a node cluster, then the node is more representative of the node cluster than other nodes in the node cluster. In the scenario where a node represents a first multimedia resource in the embodiment of the present disclosure, if a node is the center of mass of a node cluster, then the first multimedia resource represented by the node is more representative of the categories corresponding to these first multimedia resources that are similar to the first multimedia resource, and this first multimedia resource is more able to reflect the core interests of the user.

[0362] After the first graph is generated, the connections between the first multimedia resources are mapped to connections between nodes. The similarity between each first multimedia resource and a candidate multimedia resource is mapped to the similarity between each node and the candidate multimedia resource in the first graph. The second multimedia resource of each first multimedia resource is mapped to the neighboring nodes of each node in the first graph. The association information between the first multimedia resource and the second multimedia resource is mapped to the weight of the first edge, where the first edge is an edge from a neighboring node to the node.

[0363] When analyzing each edge in the graph, the electronic device may refer to one of the two nodes connecting an edge as a source node and the other as a destination node. Here, the edge is directed, with the direction of the edge pointing from the source node to the destination node. This edge is referred to as a first edge below. The destination node is the aforementioned node, and the source node is a neighbor node of the node. Accordingly, in step S63, for each first edge in the first graph, the electronic device obtains the weight of the edge pointing from the source node to the destination node based on the first node parameter of the destination node, the source node, and the similarity between the destination node and the candidate multimedia resource.

[0364] In the disclosed embodiment, the process of obtaining the weight of the first edge can use an attention mechanism to analyze which node or edge requires more attention, thereby obtaining the weight of the first edge. The weight can be understood as an attention score, which is used to indicate the importance of the first edge.

[0365] Specifically, step S63 can be implemented through the following steps 1 to 3.

[0366] Step 1: The electronic device obtains a candidate weight of the first edge according to the first node parameter of each node in the first graph and the first similarity between the neighboring node of the node and the candidate multimedia resource.

[0367] Specifically, when determining the candidate weight of the first edge, the possibility of the node being the centroid of the node cluster can be evaluated. If the node is the centroid of the node cluster, then the node is the core interest node of the user account, which is naturally more important in the subsequent prediction process. More attention needs to be paid to the first edge, and the candidate weight of the first edge can be relatively large. Considering that the embodiment of the present disclosure is intended to analyze the possibility of the user account interacting with the candidate multimedia resource, when analyzing whether the neighboring node is important to the node, it can be focused on whether the neighboring node is related to the candidate multimedia resource. If so, the neighboring node is more important, and more information of the neighboring node needs to be passed to the node, so that the representation information of the node is more relevant to the candidate multimedia resource, and vice versa. Therefore, when determining the candidate weight of the first edge, the first node parameter of the node and the similarity between the neighboring node and the candidate multimedia resource can be determined first, and then the two can be combined to obtain the candidate weight of the first edge.

[0368] For the first node parameter of a node, the node cluster corresponding to the node can be determined first, and then the average representation information of the node cluster can be obtained based on the representation information of each node in the node cluster. The third similarity between the representation information of the node and the average representation information is obtained, and the third similarity is used as the first node parameter of the node.

[0369] By considering multiple nodes around a node as a node cluster, and by comparing the representation information of the node with the average representation information of the node cluster, it is measured whether the node is the center of mass of the node cluster, where the center of mass refers to the mass center. It can be understood that if the nodes in the node cluster are relatively similar and the current node is the center of the node cluster, the representation information of the node will be very close to the average representation information of the node cluster. If the two are very different, it proves that the node is very far away from the center of mass of the node cluster, and naturally it is unlikely to be the center of mass of the node cluster. By comparing the representation information of the node with the average representation information of the node cluster, the possibility of the node being the center of mass of the node cluster is quantified and accurately represented by the first node parameter, which provides a data basis for analyzing the importance of the node, thereby making the core interest analysis of the user account more accurate.

[0370] Here we can assume that node v i The neighborhood of will form a cluster and the node v in the graph i Considered as cluster c(v i ) and then the node v i The k-hop neighborhood of the cluster is defined as the receptive field of the cluster. Wherein, k is a positive integer and can be set by relevant technical personnel according to needs. The embodiment of this disclosure does not limit this. The average value of the representation information of all nodes in the cluster is Represents the average representation information of the cluster. To determine whether a node is the core of a cluster, the node representation information and the average representation information of its corresponding cluster are used to calculate the possibility that the node is the center of mass of the node cluster. This possibility can be understood as the attention score of the node.

[0371] In some embodiments, the attention score can be calculated by an attention network, which can be a neural network. The attention score can be calculated by the above formula 5. Accordingly, the definitions of some parameters in the above formula 5 can be: α i is node v i The attention score of node v i The first node parameter. is node v i The representation information may be an embedding vector. is node v i The average representation information of the corresponding node cluster. is the average representation information of the node cluster.

[0372] In a specific possible embodiment, the neural network may include two fully connected layers and an activation layer. The activation layer may use any activation function. For example, in a specific example, the attention network ATTE c It is a two-layer feedforward neural network with LeakyReLU as the activation function.

[0373] By analyzing the similarity between the neighbor node and the candidate multimedia resource, the candidate multimedia resource can be used to dynamically move the relevant information in the first graph closer to the candidate multimedia resource, and the correlation between each node and the candidate multimedia resource can be analyzed in a targeted manner, so that the updated graph can more accurately determine whether the candidate multimedia resource is related to the core interests of the user.

[0374] In some embodiments, the similarity can be measured by the similarity between the characterization information. Specifically, the electronic device can consider the characterization information of the neighboring nodes. and representation information of candidate multimedia resources The correlation between them is also called similarity. If the correlation between the neighbor node and the candidate multimedia resource is higher, its weight on the node will be more obvious in the information aggregation process, and vice versa. This is also because the first multimedia resource related to the candidate multimedia resource (represented as a node here) can play a role in the final prediction process, retaining relevant information and discarding irrelevant information, so that the prediction result is not affected by irrelevant information and is more accurate.

[0375] Just as the first node parameter of the above node can be understood as the node's attention score, the similarity between the neighbor node and the candidate multimedia resource can be understood as the neighbor node's attention score, which can be calculated by an attention network, which can be a neural network. The attention score can be calculated by the above formula 6. Accordingly, the definitions of some parameters in the above formula 6 can be: β j is the neighbor node v j Attention score. is the neighbor node v j The representation information may be an embedding vector.

[0376] Similarly, the attention network ATTE q It can include two fully connected layers and an activation layer. The activation layer can use any activation function. For example, in a specific example, the attention network ATTE q It is a two-layer feedforward neural network with LeakyReLU as the activation function.

[0377] When determining the candidate weight for the first edge, two factors can be comprehensively considered: the first node parameter of the node and the similarity between the neighboring node and the candidate multimedia resource. For example, the first node parameter and the similarity can be weighted. This weighting method can include summing or averaging. Of course, other methods can also be used to combine the two factors, such as multiplication, which is not limited in the present embodiment.

[0378] In one possible implementation, the electronic device can sum the first node parameter of the node and the similarity between the neighbor node and the candidate multimedia resource to obtain the candidate weight of the edge of the neighbor node pointing to the node. Among them, the first node parameter of the node can be understood as the attention score of the node, which can also be called the cluster score. The similarity between the neighbor node and the candidate multimedia resource can be understood as the attention score of the neighbor node. Since it considers the relevance with the candidate multimedia resource, the candidate multimedia resource can be a multimedia resource to be recommended or a multimedia resource to be queried. Therefore, the attention score of the neighbor node can also be called the query score of the neighbor node. The two scores are added together, which can be used as the candidate weight of the edge of the neighbor node j to the node i.

[0379] In one specific embodiment, a node may have edges with multiple nodes. That is, a node may have multiple neighboring nodes. The similarities obtained for different neighboring nodes may vary, and may even differ significantly. When combining the two factors, the electronic device may also normalize the neighboring nodes corresponding to a node to make the contributions of different neighboring nodes comparable.

[0380] Specifically, the softmax function can be used to normalize all the choices of neighbor node j through the above formula 7. Accordingly, the definitions of some parameters in the above formula 7 can be: ij Refers to the weight of the edge from node i to node j, which is also the candidate weight of the first edge mentioned above, and can be understood as the attention score of the first edge. When j == i, the self-loop weight is obtained, and the neighborhood of node i Including node i itself, is the neighborhood of node i, including node i and node j connected to node i. Under the condition of self-loop propagation, α i Controls how much information the node can receive, and β j Controls how much information a neighbor node can send.

[0381] Step 2: The electronic device determines a second node parameter of the node and a second similarity between the node and the candidate multimedia resource according to the candidate weight of the first edge. The second node parameter is a node parameter obtained by updating the first node parameter.

[0382] The electronic device determines the candidate weight of the first edge, which can represent the information transmission from the neighboring node to the node. According to the candidate weight of the first edge, the representation information of the node can also be updated, so that the representation information of the node includes the information transmission situation, so that the representation information of the node that is more relevant to the candidate multimedia resource can be more reflected, and the representation information of similar nodes can also be made more similar.

[0383] In this way, the weight of the first edge in the graph and the node representation information are updated. The updated representation information is referred to as the target representation information. The electronic device can then determine the node that represents the user's core interest, namely the core interest node, which is referred to as the target node, and obtain the target representation information of the target node.

[0384] The electronic device can update the representation information of the node based on the candidate weight of the first edge to obtain the target representation information of the node, and then update the first node parameter of the node based on the target representation information of the node to obtain the second node parameter of the node, and determine the second similarity between the node and the candidate multimedia resource based on the target representation information of the node.

[0385] When determining the second node parameter and second similarity of a node based on the candidate weight of the first edge, the candidate weight of the first edge is determined. Naturally, based on the candidate weight, the node's representation information can be updated so that the node's representation information more clearly matches the candidate weight, resulting in more accurate representation information. When the node's representation information changes, the parameters related to the node will naturally change accordingly. By updating the relevant parameters of the node through the determined candidate weight, and then performing subsequent calculation steps based on the updated parameters, it can be ensured that the subsequent calculation process can fully consider the distribution of each node in the first graph and the relationship between each node, resulting in more accurate results.

[0386] When the nodes connected by the first edge are referred to as source nodes and destination nodes, the process may be as follows: for a destination node, the electronic device may obtain the target representation information of the destination node based on the weight of at least one first edge pointing to the destination node, the representation information of the source node of the at least one first edge, and the representation information of the destination node. The electronic device may then determine the possibility of each destination node being the target node based on the target representation information of each destination node, and then determine the destination node that meets the conditions based on the possibility as the target node, and use the target representation information of the destination node as the target representation information of the target node. The electronic device weights the representation information of the source node by the candidate weight of the first edge of each source node, so that the information of source nodes with different weights has different effects on the destination node.

[0387] During the process of acquiring the representation information of the node, it is assumed that the node set in the first graph is Where n is the number of nodes (i.e., the length of the historical interaction sequence of the user account), and d is the dimension of the representation information of each node. The representation information of the above nodes is the original representation information. Here, the updated target representation information is represented as The dimension of the target representation information is d′, where the dimension of the target representation information may be the same as or different from the dimension of the representation information.

[0388] The candidate weight e of the first edge above ij Can map out node v i At its neighbor node v j The importance of the source node, where the normalized candidate weight e ij It is used to calculate the linear combination of the representation information of the node to obtain the final representation information of each node, which is also the target representation information.

[0389] In a specific possible embodiment, the target representation information of the node can be calculated using the above formula 8. Accordingly, the definitions of some parameters in the above formula 8 can be: is the target representation information of the node, It is the representation information of the node. It is the neighborhood of a node, that is, the set of neighboring nodes of the node.

[0390] In some embodiments, the target representation can also be obtained by processing the representation information from multiple different perspectives. For example, using a summation function as the aggregation function and using φ independent perspectives, φ independent attention mechanisms perform the aforementioned transformations and then concatenate their representations, producing the following output embedding. This target representation can also be achieved using Formula 9 above.

[0391] Step 3: The electronic device determines the weight of the first edge according to the second node parameter of the node, the second similarity, the first target value, and the candidate weight of the first edge.

[0392] In the above information aggregation process, the fusion of implicit interest signals to explicit interest signals is completed. This can be understood as a graph pooling process, which aims to reasonably reduce the size of the graph. The electronic device can use the graph pooling method to further extract the fused information on the interest graph (first graph). Through the coarsening and collapse of the constructed graph structure, loose interests are transformed into tight interests. That is, the electronic device can extract the core interest node (that is, the target node, which is used to represent the above-mentioned third multimedia resource) from multiple nodes according to the importance of each node, and perform the subsequent prediction process through the core interest node.

[0393] The electronic device aggregates information on the graph through the connections between nodes and the similarities between nodes and candidate multimedia resources, so that the node parameters of the nodes after information aggregation and the similarities between the nodes and candidate multimedia resources are more accurate, which can help analyze which nodes are related to and more important to the candidate multimedia resources. Through information aggregation, subsequent predictions can be made more accurate, thereby improving the success rate and conversion rate of recommendations.

[0394] When the electronic device determines the possibility that each node is a target node, the possibility may be determined according to the importance of the node, and the importance of the node may be represented by importance indication information of the node.

[0395] In some embodiments, the electronic device may determine the importance indication information of the node based on the second node parameter of the node and the second similarity. Then, the weight of the first edge is determined based on the magnitude relationship between the importance indication information and the target value. Specifically, the electronic device may use the candidate weight of the first edge as the weight of the first edge in response to the importance indication information of the node being greater than or equal to the first target value. The electronic device may set the weight of the first edge to zero in response to the importance indication information of the node being less than the first target value.

[0396] By updating the weights of the nodes and the first edges in the first graph, the data in the updated first graph can better reflect the connection between the nodes. Based on this, it is possible to analyze whether each node is more important and whether it can play an important role in analyzing whether the candidate multimedia resources are in line with the core interests of the user account. Based on this, the weight of the first edge is determined, and the unimportant first edges in the first graph are filtered out, retaining only the most important first edges. In this way, the target nodes that can better reflect the core interests of the user account can be found, thereby ensuring the accuracy of subsequent recommendation results.

[0397] In some embodiments, the importance indication information of the node can also be implemented using an attention mechanism. After obtaining the importance indication information of the node, the weight of the first edge can be determined to evaluate the possibility that the node is the target node.

[0398] In a specific possible embodiment, the electronic device can determine the possibility γ that the node is the target node by using the above formulas 10, 11 and 12: i , which is the importance indication information of the node. Accordingly, the definitions of some parameters in the above three formulas can be: is node v i The attention score of node v i The second node parameter. refers to node v i The second similarity with the candidate multimedia resource. γ i is the node v i The importance of the information. is node v i The target representation information. c and W q is a weight vector, which may be the same as or different from the weight vector in the above-mentioned first node parameter and similarity acquisition process. It is the average representation information calculated based on the target representation information of each node in the node cluster.

[0399] The attention mechanism's parameters are identical to those in the interest fusion layer, except that the input becomes aggregated cluster information. That is, Formulas 10 through 12 are similar to Formulas 5 through 8 above, except that Formulas 5 through 8 determine the relevant parameters of nodes i and j to determine the weight of the first edge from node j to node i. Formulas 10 through 12 determine the importance indicator for node i, replacing the parameters of node j with those of node i.

[0400] After determining the probability that the node is the target node γ i After that, the node cluster corresponding to the node can be sampled according to the possibility, and the node on the centroid of the target number of clusters can be used as the target node.

[0401] The process of determining the target node can be understood as a graph pooling process, and the target number can be the pooling length or the minimum value between the pooling length and the number of nodes in the graph.

[0402] The first target value is the value in the importance indication information for each node that is greater than the second target value. The second target value is the minimum of the target length and the number of nodes in the first graph. By using a relative ranking strategy, a few representative nodes of core interest are selected, rather than using all nodes for prediction. This reduces computational complexity and produces prediction results that are more consistent with the user's core interests.

[0403] Assuming a given pooling length m < n, select the m clusters with the highest scores and include them in the pooled graph. In order to obtain the pooled graph, a cluster assignment matrix E is necessary. Since the attention score e in the interest fusion layer ij Considering the affiliation between nodes and clusters, this matrix is ​​reused as the cluster assignment matrix. Considering the above two points at the same time, the cluster assignment matrix after sampling is It can be obtained by the above formula 12.

[0404] After extracting the important interests, the original adjacency matrix A and the sampled cluster assignment matrix According to the above formula 13, the pooled graph is Generate a new adjacency matrix To ensure the connectivity strength between each pair of clusters. In Formula 13, if there is any common node in the original graph G or any of its constituent nodes are neighbors, then Any two clusters c(i) and c(j) in are connected. Therefore, the strength of the connection between clusters is determined by the membership of the constituent nodes. and edge weight A is determined.

[0405] Through the above steps, a dense, coarsened graph representing a stronger interest signal for the user account is obtained. All nodes in the dense, coarsened graph are target nodes, which are used to represent the third multimedia resource. The dense, coarsened graph is also a set corresponding to the third multimedia resource.

[0406] Step S64: The electronic device determines, from the first graph, a plurality of target nodes corresponding to the candidate multimedia resource according to the weight of each first edge in the first graph, where the weight of the first edge connecting the plurality of target nodes is not zero.

[0407] Steps S63 and S64 are processes for determining multiple target nodes corresponding to each candidate multimedia resource from the first graph based on the first node parameter of each node in the first graph, the similarity between the node and its neighboring nodes and multiple candidate multimedia resources, wherein the first node parameter is used to indicate the possibility that the node is the centroid of the node cluster, the similarity between the third multimedia resource represented by the target node and the candidate multimedia resource satisfies the similarity condition, and the first node parameter of the target node satisfies the parameter condition. That is, the target node is the aforementioned core interest node.

[0408] The above only uses the two factors of the first node parameter and the similarity between the node and the candidate multimedia resource to first determine the weight of the first edge, and then determine the target node based on the weight of the first edge as an example. By determining the weight of the first edge through two factors, the unimportant first edges in the first graph can be filtered out through these two factors, and the important first edges can be retained. In this way, considering that the nodes connected by the important first edges are more important, the target nodes can be accurately screened out, so that more accurate recommendation results can be obtained for multimedia resource recommendation.

[0409] In some embodiments, the electronic device may not determine the weight of the first edge, but may directly obtain the importance indication information of each node and filter out the target node based on the relationship between the importance indication information and the importance threshold. Of course, other methods may also be used, and the embodiments of the present disclosure are not limited thereto.

[0410] The weights of all edges are determined in the above manner. The weights of the edges connected to the core interest nodes are non-zero, while the weights of other edges are zero. The electronic device can select the core interest node based on the weight of the edge.

[0411] In some embodiments, step S63 and step S64 can be implemented through a graph convolutional network. The electronic device inputs the first graph and multiple candidate multimedia resources into the graph convolutional network, and the graph convolutional network performs convolution processing and pooling processing on the first graph based on each candidate multimedia resource, and outputs multiple target nodes corresponding to each candidate multimedia resource.

[0412] Step S65 : The electronic device determines a target multimedia resource from the multiple candidate multimedia resources according to the multiple target nodes and the interaction time sequence of the multiple target nodes corresponding to the third multimedia resource.

[0413] After the electronic device obtains the core interest node (target node) of the user account, it can use the target node as a basis for predicting and recommending multimedia resources to perform a prediction step, which is a step of determining the target multimedia resource based on the target node.

[0414] Considering that the interests of the user account may change over time, the changes in the interests of the user account over time can be analyzed by combining the first multimedia resource that has been interacted with and the interaction time.

[0415] The electronic device can make recommendations based on the interaction time sequence of the third multimedia resource represented by each target node. In a manner in which the first multimedia resource that has been interacted with is converted into a historical interaction sequence, the interaction time sequence of the first multimedia resource corresponding to each node is also the relative position in the historical behavior sequence. The electronic device can flatten the fused interest nodes in the pooled graph. Each element in the flattened sequence represents an enhanced interest signal.

[0416] Specifically, the electronic device can obtain an interaction sequence for each candidate multimedia resource, where the interaction sequence is composed of the multiple target nodes, and the multiple target nodes are arranged in chronological order according to the interaction time of the corresponding third multimedia resource. The electronic device can obtain a matching degree between the interaction sequence of each candidate multimedia resource and each candidate multimedia resource, and use the candidate multimedia resource with the greatest matching degree as the target multimedia resource. Taking into account the chronological order of the user account's interaction with the third multimedia resource, the interaction sequence can be used to indicate both the core interests of the user account and changes in the core interests of the user account. This allows for rapid and accurate determination of whether the candidate multimedia resource matches the core interests of the user account and changes in the core interests, thereby obtaining accurate recommendation results.

[0417] The matching degree acquisition process can be achieved through the following steps.

[0418] Step 1: For each candidate multimedia resource, the electronic device obtains first representation information of a second graph according to the multiple target nodes, where the second graph is composed of the multiple target nodes.

[0419] The core interests of the user are summarized by the first representation information to obtain the core interest expression of the user. Specifically, for the graph composed of the target nodes, the interest features of the user account can be obtained by weighting the representation information of the target nodes. The interest features can be represented at the graph level. That is the first characterization information, which can be obtained by the above formula 14. The level representation of this figure This refers to the set level representation in the above step S44, except that the set is currently represented in the form of the second graph.

[0420] Step 2: The electronic device obtains second representation information according to the importance indication information of each target node and the change of the target node in the interaction sequence, and the second representation information conforms to the change.

[0421] Any known sequential recommendation method can be used to model the node sequence consisting of target nodes. Assuming a single sequential recommendation model is used to model the evolution of interests, the evolution method can be as shown in Formula 15 above, evolving the core interest changes of the user account through m target nodes. Accordingly, the definitions of some parameters in Formula 15 above can be: It is the evolution result of the sequential recommendation model on the node sequence, which is also the second representation information.

[0422] In some embodiments, the above-mentioned second representation information can be implemented through an interest evolution model. The electronic device can input the importance indication information of each target node and the interaction sequence into the interest evolution model and output the second representation information. The AUGRU is the interest evolution model.

[0423] Step 3: The electronic device obtains a matching degree between the candidate multimedia resource and the corresponding interaction sequence according to the first representation information, the second representation information, and the representation information of the candidate multimedia resource.

[0424] After obtaining the core interest expression, the representation information of the next interactive multimedia resource is also evolved. Through the above two representation information and the representation information of the candidate interest, it can be determined whether the candidate multimedia resource meets the requirements.

[0425] By combining the first representation information of the second figure, the second representation information that conforms to the changes in core interests, and the representation information of the candidate multimedia resources, it is possible to comprehensively and accurately analyze whether the candidate multimedia resources conform to the user's core interests and the changes in core interests, providing a data basis for multimedia resource recommendations.

[0426] The graph-level representations of the interest extraction layer and the evolutionary outputs of the interest evolution layer are used as the user's current interests and are concatenated with the representations of candidate multimedia resources. Given the concatenated dense representation vectors, a fully connected layer is used to automatically learn the combination of representations. A two-layer feedforward neural network is used as the prediction function Pred to estimate the likelihood of the user interacting with the candidate multimedia resources next time. The multimedia resource with the highest probability is then selected as the target multimedia resource.

[0427] In some embodiments, the electronic device can obtain the Hadamard product of the first representation information and the representation information of the candidate multimedia resource, and then connect the first representation information, the second representation information, the representation information of the candidate multimedia resource and the Hadamard product to obtain third representation information, and then perform dimensionality reduction processing on the third representation information to obtain the matching degree between the interaction sequence of the candidate multimedia resource and the candidate multimedia resource.

[0428] By obtaining the Hadamard product, the first representation information and the representation information of the candidate multimedia resources can be completely retained, and the connection between the current core interest node and the candidate multimedia resources can be reflected, and then the third representation information is obtained by splicing with the first representation information, the second representation information and the representation information of the candidate multimedia resources. In this way, the third representation information includes the graph, the candidate multimedia resources, the connection between the graph and the candidate multimedia resources, and the changes in the user's core interests. Based on this, dimensionality reduction processing is performed, and this information is comprehensively considered to obtain a matching degree that comprehensively considers multiple angles. The matching degree is more accurate, and multimedia resource recommendations can be made in this way to improve the accuracy of the recommendation results.

[0429] The dimensionality reduction process can be implemented in various ways. In some embodiments, the dimensionality reduction process can be implemented by a multimedia resource prediction model, where the electronic device inputs the third representation information into the multimedia resource prediction model and outputs a matching degree between the interaction sequence of the candidate multimedia resource and the candidate multimedia resource.

[0430] For example, the dimensionality reduction process can be achieved through a fully connected layer. Assuming that the multimedia resource prediction model can be a two-layer feedforward neural network, the result of the multimedia resource prediction model can be as follows: Figure 5 As shown, the two hidden layers can be implemented using a fully connected layer. The fully connected layer is used to perform a product operation on the input data and the weight, and then sum the product with the bias to obtain the output. This operation process is similar to the convolution operation performed by the 1X1 convolution layer. Therefore, the above-mentioned multimedia resource prediction model can also be implemented by a convolution layer, which is not limited in this embodiment of the present disclosure. The above-mentioned process of determining the target multimedia resource can be implemented by the above-mentioned formula 16.

[0431] It should be noted that the above-mentioned determination of the target node, determination of the second representation information, and acquisition of the matching degree can all be achieved through a network or model. Through the training process of the network or model, the network or model can have accurate data processing capabilities, and then the trained network or model can be used to quickly and accurately process relevant data to improve data processing efficiency and accuracy, thereby improving recommendation efficiency and accuracy.

[0432] Step S66: The electronic device recommends the target multimedia resource to the user account.

[0433] After the electronic device determines the target multimedia resource, it can send the target multimedia resource to the target account to recommend the target multimedia resource to the target account. The target multimedia resource extracts the user's core interests through the graph convolution process, and the recommendation process refers to the changes in core interests over time. Therefore, the recommendation results can be more in line with the needs of the user account, the number of interactions of the user account can be increased, the conversion rate of the recommended target multimedia resources can be increased, the recommendation success rate can be increased, and the recommendation effect can be better.

[0434] Step S65 and step S66 are the process of recommending the target multimedia resource from the multiple candidate multimedia resources to the user account based on the multiple target nodes. The process takes into account the fact that the user's interests may change over time. In the prediction process, the target multimedia resource is determined in accordance with the user's interest changes by combining the interaction time sequence with the third multimedia resource. This is more in line with the current interests of the user account, and the recommendation result is more accurate.

[0435] In some embodiments, the graph convolutional network, interest evolution model, and multimedia resource prediction model are considered as a large model, which can be called SURGE (SeqUential Recommendation with Graphconvolutional nEtworks, Sequential Recommendation based on Graphconvolutional Network) model. Figure 7 As shown in the figure, based on metric learning, each interaction sequence is reconstructed into an interest graph (A), and interest fusion (B) and extraction (C) are dynamically performed on this interest graph. By expanding the coarsened graph after fusion and extraction, the currently activated core interest sequence (D) is obtained, which can be used for further modeling interest evolution and prediction.

[0436] Among them, based on metric learning, each interaction sequence is reconstructed into an interest graph (A) by Figure 6 (a) and (b) are shown in Figure 2. In (a), the first multimedia resource that the user interacts with is represented as an interaction sequence. Reconstructing it yields the interest graph in (b).

[0437] The detailed process for interest fusion (B) and extraction (C) is as follows: As shown in (c), the cluster attention score of each node is calculated. Here, the score is represented by a fill pattern; more fill patterns indicate a higher score, and high-scoring nodes represent the user's core interests. For simplicity, only three clusters are plotted. This is merely a display method; different nodes can also be colored, with the color depth used to determine the cluster attention score. As shown in (d), the query attention score of the source node (neighboring node) is calculated. The more fill patterns, the stronger the relevance to the target item (i.e., the candidate multimedia resource). As shown in (e), interest fusion is performed through attention propagation. For clarity, only nodes with high cluster attention scores are considered target nodes, i.e., information aggregation nodes for the source node. Arrows indicate information flow during aggregation, with the weight being the sum of the two attentions. As shown in (f), the evolution of core interests related to the target item is independently modeled by using the activated core interest scores as pooling criteria. As shown in (g), interest extraction is performed through graph pooling, transforming noisy and loose interests into compact ones.

[0438] Then, as shown in (h) and (i), the graph composed of target nodes (i.e., the second graph) is arranged in the order of interaction time through the graph pooling layer, and another interaction sequence is obtained by spreading it out. This interaction sequence includes the user's core interests and nodes (i.e., target nodes).

[0439] according to Figure 7 The SURGE model shown in Figure 1 consists of four parts: the interest graph construction part, the interest fusion graph convolution layer, the interest extraction graph pooling layer, and the prediction layer. This architecture will be described in detail component by component.

[0440] (1) Interest graph construction. This part constructs a compact item-item interest graph from a lengthy item sequence based on metric learning, explicitly integrating and distinguishing different types of preferences in long-term user behavior. This interest graph construction corresponds to step S62 above.

[0441] (2) Interest Fusion Graph Convolution Layer. This interest fusion graph convolution layer dynamically integrates the user's interests through graph convolution propagation on the constructed interest graph, thereby strengthening important behaviors and weakening noisy behaviors. The use of this interest fusion graph convolution layer corresponds to the above step S63.

[0442] (3) Interest extraction graph pooling layer. The interest extraction graph pooling layer takes into account the different preferences of users at different times. Further dynamic graph pooling operations can adaptively retain the currently activated preferences. The use of this interest extraction graph pooling layer corresponds to the above step S64.

[0443] (4) Prediction layer. The prediction layer flattens the pooled graph into a streamlined sequence, models the evolution of the enhanced interest signal, and predicts the user's next interaction. The steps performed by the prediction layer correspond to the above-mentioned step S65.

[0444] above Figure 4 and Figure 6 The embodiment shown has described the specific process of the multimedia resource recommendation method. The above multimedia resource recommendation method can be implemented by a graph convolutional network, an interest evolution model and the multimedia resource prediction model. Figure 8 The illustrated embodiment illustrates the training process of the graph convolutional network, the interest evolution model, and the multimedia resource prediction model.

[0445] S81. The electronic device obtains sample multimedia resources that the sample user account has interacted with.

[0446] The step S81 is similar to the above step S61 and will not be described in detail here.

[0447] S82: The electronic device uses any sample multimedia resource as a destination multimedia resource.

[0448] After obtaining the first historical behavior data of the sample user account, the multimedia resources that the sample user account has interacted with in the first historical behavior data are all known and can be split. The multimedia resources that come earlier in the interaction time sequence are used as the multimedia resources for analyzing user interests, and the multimedia resources that come later are used as the multimedia resources for judging whether to interact. The second historical behavior data obtained in this way is the multimedia resource for analyzing user interests, and the target multimedia resource is the label data of the second historical behavior data, which is used to identify the correct multimedia resource to be interacted with next time obtained by analyzing the second historical behavior data.

[0449] S83: The electronic device generates a sample first graph based on other sample multimedia resources whose interaction time is before the target multimedia resource.

[0450] The step S83 is similar to the above step S62 and will not be described in detail here.

[0451] S84. The electronic device inputs the sample first graph and multiple candidate multimedia resources into a graph convolutional network, and the graph convolutional network performs convolution processing and pooling processing on the first graph based on the multiple candidate multimedia resources to obtain multiple target nodes corresponding to each candidate multimedia resource, and the multiple candidate multimedia resources at least include the target multimedia resource.

[0452] S85. The electronic device processes the multimedia resources corresponding to the multiple target nodes corresponding to each candidate multimedia resource based on the interest evolution model and the multimedia resource prediction model, and outputs the predicted multimedia resources.

[0453] Step S84 and step S85 are similar to the above-mentioned steps S63 to S64, and will not be described in detail here.

[0454] S86. The electronic device trains the graph convolutional network, the interest evolution model, and the multimedia resource prediction model according to the difference between the predicted multimedia resource and the target multimedia resource.

[0455] The predicted multimedia resource determined in the above step S86 is the prediction result, and the target multimedia resource in the above step S82 is the correct result. By comparing the predicted result and the correct result, it is possible to predict whether the result is accurate, so as to judge the performance of the above graph convolutional network, interest evolution model and multimedia resource prediction model, and then determine whether its parameters need to be updated to optimize the model.

[0456] When comparing the two, the accuracy of the prediction can be characterized by the difference between the two, which can be achieved by the loss value determined by any loss function.

[0457] In some embodiments, the sequential recommendation process is similar to the CTR (click-through rate) estimation process, and the negative log-likelihood function (logarithmic loss) can be used as the loss function. The difference can be obtained by the following formula 17:

[0458]

[0459] in, is the training set, is the number of training instances (i.e., the second historical behavior data). Θ represents the set of trainable parameters, while λ controls the penalty intensity. Label y o =1 indicates a positive example, y o =0 indicates a negative example. is the network output after the softmax layer, representing the predicted probability of the next multimedia resource in the interaction. L is the difference between the predicted multimedia resource and the target multimedia resource, which is also the loss value.

[0460] If the loss value is greater than a threshold, the electronic device can update the parameters of the three models and then perform the next iteration based on the updated parameters. If the loss value is less than the threshold, the electronic device can use the three current models as the final trained models. The optimization process is the process of minimizing the loss function, which can prevent overfitting.

[0461] By obtaining relevant data from sample user accounts, the multimedia resources that users have interacted with in the relevant data are used to establish training instances according to the multimedia resources interacted with at different time points. By training the model with real user data, the model's data processing results can be close to the real situation, and the model's recommendation results are more accurate. In addition, based on the graph convolutional network, the user account data is converted into a graph for analysis. By analyzing the connections between nodes in the form of a graph, the user account's interests can be explicitly integrated and analyzed, and the user account's interests can be visualized. Subsequently, based on the first node parameters of the nodes in the graph, it is possible to analyze which nodes are the core interests of the user account based on the distribution of the nodes in the graph. The similarity between the nodes and the candidate multimedia resources can be used to analyze which nodes are similar to the candidate multimedia resources. Combining these two factors, it is natural to analyze multimedia resources that are similar to the candidate multimedia resources and can represent the core interests of the user account from the multimedia resources that the user account has interacted with. In this way, multimedia resources can be recommended for the user account, ensuring that the recommended target multimedia resources are consistent with the core interests of the user account and meet the user's preferences, thereby improving the accuracy of the recommendation results.

[0462] The following is an experimental comparison of the technical solutions provided by the present disclosure and the technical solutions provided by related technologies. In the experiment, extensive experiments were carried out on public data sets and industry data sets to evaluate the technical solutions of the present disclosure. The experimental results show that the method proposed in the present disclosure has significant performance improvements compared with the methods provided by several related technologies. Further research has confirmed that the method provided by the present disclosure can effectively and efficiently model longer behavior sequences. In the embodiments of the present disclosure, the above-mentioned multimedia resources are referred to as items, which will be explained below through experimental settings, the overall performance of each technical solution, and sequence length research and efficiency comparison.

[0463] 1. Experimental Setup

[0464] To demonstrate the effectiveness of the proposed SURGE model (the solution provided by this disclosure), we compared it with sequential recommenders in related technologies. The baselines are divided into two categories: non-sequential recommendation models that only capture users' static interests, and sequential recommendation models that consider dynamic interest patterns.

[0465] 1) Dataset

[0466] The recommendation performance is evaluated on a public e-commerce dataset and an industry-specific short video dataset. Table 1 summarizes the basic statistics of these two datasets. The average length represents the average length of user history records, indicating that the scale of the industry dataset used is much larger than the public dataset.

[0467] Table 1

[0468] Dataset user thing Instances Average length Dataset 1 36915 64138 1471155 39.85 Dataset 2 60813 292286 14952659 245.88

[0469] Dataset 1: This dataset, collected from China's largest e-commerce platform, is widely used in recommendation research. The experiments used click data from November 25 to December 3, 2017, filtering out users with fewer than 10 interactions. The first seven days served as the training set, the eighth day as the validation set, and the last day as the test set.

[0470] Dataset 2: This is an industry-specific dataset collected from one of the largest short video platforms. Users can upload short videos and browse other users' short videos. The data in the dataset is obtained by downsampling logs from October 22 to October 28, 2020. The dataset records various user behaviors (such as clicks, likes, follows, and reposts). Click data is used in the experiment, and a 10-core configuration is also used to filter out invalid samples. The samples from the first 6 days are used to train the recommendation model. The samples before 12:00 noon on the last day are used as the validation set, and the samples after 12:00 noon are used to evaluate the final recommendation performance.

[0471] 2) Evaluation Metrics

[0472] To evaluate the performance of each model, we used two widely adopted accuracy metrics: Area Under the Curve (AUC) and Group Area Under the Curve (GAUC), as well as two ranking metrics: Mean Reciprocal Rank (MRR) and Normalized Discounted Cumulative Gain (NDCG). AUC represents the probability that a sample with user interaction has a higher score than a sample with user interaction, reflecting the classification model's ability to rank samples. GAUC takes a weighted average of each user's AUC, where the weight is the number of clicks. This eliminates the impact of user bias on the model and provides a more granular assessment of model performance. MRR is the mean reciprocal rank, which is the average of the last ranked items in the sequence for the first hit. NDCG@K assigns higher scores to hits that are ranked higher in the top K ranking list, emphasizing that test items should be ranked as high as possible. In our experiments, K was set to 2, a setting widely used in existing work.

[0473] 3) Baseline

[0474] To demonstrate the effectiveness of the SURGE model, we compared it with several sequential recommenders from related technologies. The baselines fall into two categories: non-sequential recommendation models that only capture static user interests, and sequential recommendation models that consider dynamic interest patterns. The baselines refer to sequential recommenders from related technologies. Below is a brief introduction to the recommendation models from related technologies used for comparison in this disclosure.

[0475] Non-sequential recommendation models include DIN. DIN uses an attention mechanism and takes the target item as a query vector. It obtains a user representation by aggregating historical interaction sequences with attention weights. Sequential recommendation models include DIEN, Caser, GRU4REC, and SLi-Rec. DIEN uses a two-layer GRU consisting of an interest extraction layer and an interest evolution layer to model user behavior sequences. Caser embeds item sequences that are close in time and latent space into image features and uses convolutional kernels to learn their sequential patterns. GRU4REC uses GRU to model user session sequences and encodes user interests into the final state. SLi-Rec is a state-of-the-art method for jointly modeling long-term and short-term interests, based on an attention framework and an improved time-aware LSTM.

[0476] 4) Hyperparameter setting

[0477] The SUGER model provided in the present disclosure can implement all models using the Microsoft Recommenders framework based on TensorFlow. Among them, TensorFlow is a symbolic mathematical system based on data flow programming. Then, the Adam (adaptive moment estimation) algorithm is used for optimization during the model training process, with an initial learning rate of 0.001. For all models, the batch size is set to 500 and the dimension of the representation information is fixed to 40. Then, Xavier initialization is used to initialize the parameters. All methods use a two-layer feedforward neural network with a latent vector size of [100, 64] for interaction estimation. The maximum length of the user interaction sequence is 50 on dataset 1 and 250 on dataset 2. Grid search is applied to find the optimal hyperparameters. Specifically, the L2 regularization coefficient is searched in [1e-7, 1e-6, 1e-5]; the pooling length of the user interaction sequence of dataset 1 and dataset 2 is searched in [10, 20, 30, 40, 50] and [50, 100, 150, 200, 250] respectively.

[0478] 2. Overall Performance

[0479] Through experiments, Table 2 shows the results on two datasets. From the results, we can make the following observations:

[0480] The SURGE model achieves the best performance. It can be observed that the SURGE model significantly outperforms all baselines in both classification and ranking metrics. Specifically, the SURGE model improves the AUC by approximately 0.03 (p-value < 0.001) on Dataset 1 and by nearly 0.04 (p-value < 0.001) on Dataset 2. The improvement is even more pronounced on Dataset 2, which has a longer interaction history, indicating that the pooling method can more effectively handle long sequences and significantly reduces the difficulty of modeling user interests.

[0481] Sequential recommendation models in the related art are effective but suffer from short-term memory bottlenecks. The superior performance of Caser, DIEN, and GRU4Rec compared to NCF, DIN, and LightGCN demonstrates the necessity of capturing sequential patterns for modeling user interests. On Dataset 1, RNN-based models (GRU4Rec and DIEN), which have a stronger ability to capture sequential patterns, outperformed the CNN-based model (Caser). This is because max pooling in CNNs, a common and effective approach in computer vision, ignores important positional and recurrent signals when modeling long-range sequential data. However, on Dataset 2, because RNNs tend to forget long-term interests when processing longer sequences, the performance of DIEN and GRU4Rec is matched or even surpassed by Caser in certain metrics. This result demonstrates that even powerful recurrent neural networks have short-term memory bottlenecks. Furthermore, because long sequences tend to contain more noise, DIEN's performance on both datasets is unstable compared to GRU4REC. This shows that the two-layer GRU (Gated recurrent unit) structure, which should be more effective, is more easily interfered by noise on datasets with longer sequences, thus proving that the motivation of summarizing sequences is reasonable.

[0482] In addition, the joint modeling of long-term and short-term interests does not always lead to better performance. Sli-Rec, a joint model of long-term and short-term interests, is the best baseline on dataset 1 in terms of the AUC metric, but performs poorly according to ranking metrics. Moreover, on dataset 2 where the interaction sequences are longer, Sli-Rec performs worse than GRU4Rec in all metrics, even though GRU4REC does not explicitly distinguish between long-term and short-term interests. This shows that although Sli-Rec utilizes two independent components to model users' long-term and short-term interests, it still cannot effectively integrate them into a single model, especially for the processing of long sequences. In addition, Sli-Rec utilizes timestamp information to improve modeling long-term and short-term interests. However, by using metric learning to compress information, methods can show better performance without explicitly modeling interaction time.

[0483] Table 2

[0484]

[0485]

[0486] 3. Sequence Length Study and Efficiency Comparison

[0487] The following sequence length study and efficiency comparison of each method are conducted to explore whether the method proposed in this disclosure is more effective and efficient in processing long sequences. First, users are grouped and the performance of different models is tested under different sequence length groups. Then the convergence process of the model during training is visualized and the training efficiency of different models is compared. For clarity, the most competitive sequential method is plotted: for non-sequential methods, only the best method DIN is plotted (i.e., excluding NCF and LightGCN).

[0488] 1) Sequence length study.

[0489] Users may have very long interaction sequences. Longer historical sequences usually have more patterns that reflect user interests, but the increased noise signals will mislead the modeling of real interests. Therefore, whether to effectively model the long-term history of users is an important issue for sequential recommendation. Specifically, in the experiment, all users of the two datasets are divided into five groups according to the length of the interaction history. For each group, the performance of the method provided by the present disclosure is compared with the baseline method, and the GAUC performance indicators of the two datasets are given, such as Figure 9 shown.

[0490] The results show that when sequence length is short, all models struggle to capture users' true interests due to data sparsity. As sequence length increases and modeling difficulty decreases, the performance of most models improves and reaches a peak. However, as sequence length continues to increase, the performance of almost all models declines due to the introduction of significant noise. DIN and DIEN experience the largest performance drops. DIN struggles to focus on the most critical elements in long sequences. The most popular items may appear early in the sequence and be quite distant from the user's current interests. In DIEN's two-layer GRU architecture, the input to the lower GRU depends on the output of the upper GRU when modeling user interests, making it more susceptible to noise in long sequences. GRU4REC, due to the short-term bottleneck of a single GRU, focuses only on recent history and ignores sequence length. However, its performance remains relatively stable across all length groups. Although Sli-Rec considers both long-term and short-term user interests, it also suffers from performance degradation on long sequences because it models noisy sequences.

[0491] The performance gap between SURGE (proposed in this disclosure) and other methods widens as sequence length increases. For users with the longest histories, SURGE maintains excellent performance of 0.8919 and 0.8502 on Dataset 1 and Dataset 2, respectively. This is because the SURGE model merges implicit signals into explicit signals and filters out noise, thus achieving good performance for users with longer histories. Therefore, the SURGE model proposed in this disclosure can more effectively model long-term user histories.

[0492] 2) Efficiency comparison.

[0493] For sequential recommendation systems, efficiently modeling user behavior history is a big challenge. The differences and diversity of items in user history sequences lead to slow model convergence. In addition, long history sequences often correspond to more complex calculations and more time-consuming training. In the experiment, the training process of SURGE and the baseline model was visualized, and the convergence speed and training time of each model were compared. Specifically, the performance changes of the proposed method and the baseline method on the validation set during the training process were plotted, and the GAUC indicator was reported. Early stopping is used to detect whether the training is over, that is, if the GAUC on the validation set does not increase within five periods, the training process will stop. For the performance change curves of the two data sets, smoothing rates of 0.2 and 0.6 are used to smooth them respectively to better see the trend.

[0494] The training process of the two datasets is as follows Figure 10 As shown in the figure, DIN struggles to focus on key interests in longer sequences, resulting in constant fluctuations and difficulty converging on Dataset 2. Because GRU4REC exhibits a more pronounced tendency to forget long-term patterns, only the embedding of the last item in each training instance is updated. This results in a smooth and slow training curve, with minimal growth, making it difficult to stop early. Because SLIREC specifically considers long-term user interests, it converges quickly on Dataset 2, but is the slowest method to converge on Dataset 1, which has shorter sequences.

[0495] Table 3 shows the training time of each model on the two datasets. As shown in Table 3, except for the non-sequential recommendation model DIN on dataset 2, the efficiency of the method provided by the present disclosure is improved by more than 20% compared with all baselines. This is because SURGE performs a pooling operation on the sequence before feeding the embedded sequence into the recurrent neural network, which greatly reduces the cycle step size of the recurrent neural network. In addition, since most of the noise has been filtered out, the pooled sequence only contains the most core interests, which undoubtedly helps to speed up the convergence of the model. Therefore, the SURGE model proposed in this disclosure can more efficiently model the long-term historical sequences of users.

[0496] Table 3

[0497]

[0498] Since the method provided by the present disclosure is independent of the choice of prediction layer after pooling the sequence, any known sequential recommendation method can be used to model the sequence of interest after pooling. The following compares the effects of using different prediction layers on the pooled sequence, including Attention (DIN), GRU (GRU4Rec), AUGRU (DIEN) and TIME4LSTM (Sli-Rec). The results are shown in Figure 2. Figure 11 As shown in Figure 2, the first observation is that the performance of sequential recommendation models other than DIN is not significantly different, with AUGRU, which leverages cluster scores in the interest extraction layer, performing slightly better. The second observation is that modeling on pooled sequences yields gains for all existing methods. This suggests that a pooling strategy significantly reduces the difficulty of modeling user interests and achieves better performance.

[0499] In this experiment, extensive experiments were conducted on two real data sets. The experimental results proved that the SURGE model provided by the present disclosure is superior to the multimedia resource recommendation method in the related art. Further research shows that the SURGE model can effectively and efficiently alleviate the problem that long sequences are difficult to model. The technical solution provided by the present disclosure approaches the sequential recommendation problem from a new perspective by considering implicit signal behavior and rapidly changing preferences. The technical solution proposes to aggregate the implicit signals in user behavior into explicit signals by designing a graph neural network-based model on the constructed project-item interest graph. Then, dynamic pooling is designed to filter and retain activated preferences for recommendation. In addition, the experimental results also show that the performance has been significantly improved compared with the sequential recommendation method in the related art. Further research also proves that the method provided by the present disclosure can effectively and efficiently model long behavior sequences.

[0500] Figure 12 FIG. 1 is a block diagram of a multimedia resource recommendation device according to an exemplary embodiment. Figure 12 The device includes an acquisition unit 1201, a determination unit 1202 and a recommendation unit 1203.

[0501] An acquiring unit 1201 is configured to acquire a multimedia resource set that a user account has interacted with, the multimedia resource set including a plurality of first multimedia resources;

[0502] Determining unit 1202 is configured to determine, from the multimedia resource set, a third multimedia resource corresponding to each candidate multimedia resource based on a first resource category parameter of each first multimedia resource, a similarity between the first multimedia resource and a second multimedia resource of the first multimedia resource and a plurality of candidate multimedia resources, wherein the first resource category parameter is used to indicate a likelihood that the first multimedia resource can represent a corresponding category, the second multimedia resource is a multimedia resource in the multimedia resource set whose similarity to the first multimedia resource satisfies a first similarity condition, the similarity between the third multimedia resource and the candidate multimedia resource satisfies a second similarity condition, and the first resource category parameter of the third multimedia resource satisfies a parameter condition;

[0503] The recommendation unit 1203 is configured to perform multimedia resource recommendation to the user account based on the multiple third multimedia resources.

[0504] In some embodiments, the determining unit 1202 is configured to perform:

[0505] For each candidate multimedia resource among the plurality of candidate multimedia resources, determining association information between the first multimedia resource and the second multimedia resource based on the first resource category parameter of each first multimedia resource, the first multimedia resource, and a similarity between the second multimedia resource of the first multimedia resource and the candidate multimedia resource;

[0506] According to the association information between the first multimedia resource and the second multimedia resource, a plurality of third multimedia resources corresponding to the candidate multimedia resource are determined from the multimedia resource set, and the association information corresponding to the plurality of third multimedia resources is not zero.

[0507] In some embodiments, the determining unit 1202 is configured to perform:

[0508] Acquire candidate association information between the first multimedia resource and the second multimedia resource according to the first resource category parameter of each first multimedia resource and the first similarity between the second multimedia resource of the first multimedia resource and the candidate multimedia resource;

[0509] Determining, based on the candidate association information between the first multimedia resource and the second multimedia resource, a second resource category parameter of the first multimedia resource and a second similarity between the first multimedia resource and the candidate multimedia resource, wherein the second resource category parameter is a resource category parameter obtained by updating the first resource category parameter;

[0510] The association information between the first multimedia resource and the second multimedia resource is determined according to the second resource category parameter of the first multimedia resource, the second similarity, the first target value, and the candidate association information between the first multimedia resource and the second multimedia resource.

[0511] In some embodiments, the determining unit 1202 is configured to perform:

[0512] Based on the candidate association information between the first multimedia resource and the second multimedia resource, updating the representation information of the first multimedia resource to obtain target representation information of the first multimedia resource;

[0513] Based on the target representation information of the first multimedia resource, a first resource category parameter of the first multimedia resource is updated to obtain a second resource category parameter of the first multimedia resource;

[0514] Based on the target representation information of the first multimedia resource, a second similarity between the first multimedia resource and the candidate multimedia resource is determined.

[0515] In some embodiments, the determining unit 1202 is configured to perform:

[0516] determining importance indication information of the first multimedia resource according to the second resource category parameter of the first multimedia resource and the second similarity;

[0517] In response to the importance indication information of the first multimedia resource being greater than or equal to a first target value, using the candidate association information between the first multimedia resource and the second multimedia resource as the association information between the first multimedia resource and the second multimedia resource;

[0518] In response to the importance indication information of the first multimedia resource being less than a first target value, the association information between the first multimedia resource and the second multimedia resource is set to zero.

[0519] In some embodiments, the first target value is a value greater than the second target value in the importance indication information of each first multimedia resource, and the second target value is a minimum value of the target length and the number of first multimedia resources in the multimedia resource set.

[0520] In some embodiments, a method for obtaining the first resource category parameter of the first multimedia resource includes:

[0521] Determining a target category corresponding to the first multimedia resource;

[0522] Obtaining average representation information of the first multimedia resources of the target category according to the representation information of each first multimedia resource belonging to the target category;

[0523] A third similarity between the representation information of the first multimedia resource and the average representation information is obtained, and the third similarity is used as a first resource category parameter of the first multimedia resource.

[0524] In some embodiments, the recommendation unit 1203 includes a determination module and a recommendation module;

[0525] The determining module is configured to determine a target multimedia resource from the plurality of candidate multimedia resources based on the plurality of third multimedia resources and an interaction time sequence of the plurality of third multimedia resources;

[0526] The recommendation module is configured to recommend the target multimedia resource to the user account.

[0527] In some embodiments, the determination module includes an acquisition submodule and a determination submodule;

[0528] The acquisition submodule is configured to execute an interaction sequence for acquiring each candidate multimedia resource, where the interaction sequence is composed of the plurality of third multimedia resources arranged in an interaction time sequence;

[0529] The acquisition submodule is further configured to execute acquisition of a matching degree between the interaction sequence of each candidate multimedia resource and each candidate multimedia resource;

[0530] The determination submodule is configured to execute the process of using the candidate multimedia resource with the greatest matching degree as the target multimedia resource.

[0531] In some embodiments, the acquisition submodule is configured to perform:

[0532] For each candidate multimedia resource, obtaining first representation information corresponding to the plurality of third multimedia resources based on the plurality of third multimedia resources, where the first representation information is used to represent a set consisting of the plurality of third multimedia resources;

[0533] acquiring second representation information according to the importance indication information of each third multimedia resource and a change in the third multimedia resource in the interaction sequence, where the second representation information conforms to the change;

[0534] A matching degree between the interaction sequence of the candidate multimedia resource and the candidate multimedia resource is obtained according to the first representation information, the second representation information, and the representation information of the candidate multimedia resource.

[0535] In some embodiments, the acquisition submodule is configured to perform:

[0536] Obtaining a Hadamard product of the first representation information and the representation information of the candidate multimedia resource;

[0537] Connecting the first representation information, the second representation information, the representation information of the candidate multimedia resource, and the Hadamard product to obtain third representation information;

[0538] A dimensionality reduction process is performed on the third representation information to obtain a matching degree between the interaction sequence of the candidate multimedia resource and the candidate multimedia resource.

[0539] The device provided by the embodiment of the present disclosure integrates and analyzes the interests of the user account through the multimedia resources that the user account has interacted with, analyzes which multimedia resources are the core interests of the user account through the clustering of multimedia resources determined according to similarity, and analyzes which multimedia resources are similar to the candidate multimedia resources through the similarity between the multimedia resources and the candidate multimedia resources. Combining these two factors, it is naturally possible to analyze multimedia resources that are similar to the candidate multimedia resources and can represent the core interests of the user account from the multimedia resources that the user account has interacted with, so as to recommend multimedia resources to the user account, so that the recommended target multimedia resources can be consistent with the core interests of the user account and meet the user preferences, thereby improving the accuracy of the recommendation results.

[0540] Regarding the apparatus in the above embodiment, the specific manner in which each unit performs operations has been described in detail in the embodiment of the method, and will not be elaborated on here.

[0541] Figure 13 It is a structural diagram of an electronic device according to an exemplary embodiment. The electronic device 1300 may have relatively large differences due to different configurations or performances, and can include one or more processors (Central Processing Units, CPU) 1301 and one or more memories 1302, wherein the memory 1302 stores at least one instruction, and the at least one instruction is loaded and executed by the processor 1301 to implement the multimedia resource recommendation method provided by the above-mentioned various method embodiments. The electronic device can also include other components for realizing the functions of the device. For example, the electronic device can also have components such as wired or wireless network interfaces and input and output interfaces for input and output. The embodiments of the present disclosure are not described in detail here.

[0542] The electronic device in the above method embodiment can be implemented as a terminal. For example, Figure 14This is a block diagram of a terminal structure according to an exemplary embodiment. Terminal 1400 may be a portable mobile terminal, such as a smartphone, tablet computer, MP3 (Moving Picture Experts Group Audio Layer III) player, MP4 (Moving Picture Experts Group Audio Layer IV) player, laptop computer, or desktop computer. Terminal 1400 may also be referred to as user equipment, portable terminal, laptop terminal, desktop terminal, or other similar names.

[0543] Typically, the terminal 1400 includes a processor 1401 and a memory 1402 .

[0544] The processor 1401 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 1401 may be implemented in at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), or PLA (Programmable Logic Array). The processor 1401 may also include a main processor and a coprocessor. The main processor is a processor for processing data in the awake state, also known as a CPU (Central Processing Unit); the coprocessor is a low-power processor for processing data in the standby state. In some embodiments, the processor 1401 may be integrated with a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 1401 may also include an AI (Artificial Intelligence) processor, which is used to process computing operations related to machine learning.

[0545] Memory 1402 may include one or more computer-readable storage media, which may be non-transitory. Memory 1402 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices and flash memory storage devices. In some embodiments, the non-transitory computer-readable storage medium in memory 1402 is used to store at least one instruction, which is executed by processor 1401 to implement the multimedia resource recommendation method provided in the method embodiment of the present disclosure.

[0546] In some embodiments, terminal 1400 may optionally include a peripheral device interface 1403 and at least one peripheral device. Processor 1401, memory 1402, and peripheral device interface 1403 may be connected via a bus or signal lines. Each peripheral device may be connected to peripheral device interface 1403 via a bus, signal lines, or circuit boards. Specifically, the peripheral device may include at least one of a radio frequency circuit 1404, a display screen 1405, a camera assembly 1406, an audio circuit 1407, a positioning assembly 1408, and a power supply 1409.

[0547] The peripheral device interface 1403 can be used to connect at least one I / O (Input / Output)-related peripheral device to the processor 1401 and the memory 1402. In some embodiments, the processor 1401, the memory 1402, and the peripheral device interface 1403 are integrated on the same chip or circuit board; in some other embodiments, any one or two of the processor 1401, the memory 1402, and the peripheral device interface 1403 can be implemented on separate chips or circuit boards, which is not limited in this embodiment.

[0548] RF circuit 1404 is used to receive and transmit RF (Radio Frequency) signals, also known as electromagnetic signals. RF circuit 1404 communicates with communication networks and other communication devices via electromagnetic signals. RF circuit 1404 converts electrical signals into electromagnetic signals for transmission, or converts received electromagnetic signals into electrical signals. RF circuit 1404 optionally includes an antenna system, an RF transceiver, one or more amplifiers, a tuner, an oscillator, a digital signal processor, a codec chipset, a user identity module card, and the like. RF circuit 1404 can communicate with other terminals via at least one wireless communication protocol. Such wireless communication protocols include, but are not limited to, the World Wide Web, metropolitan area networks, intranets, various generations of mobile communication networks (2G, 3G, 4G, and 5G), wireless local area networks, and / or WiFi (Wireless Fidelity) networks. In some embodiments, RF circuit 1404 may also include circuitry related to Near Field Communication (NFC), although this disclosure is not limiting in this regard.

[0549] Display screen 1405 is used to display a UI (User Interface). The UI may include graphics, text, icons, videos, or any combination thereof. When display screen 1405 is a touch screen display, it is also capable of collecting touch signals on or above the surface of display screen 1405. The touch signals may be input as control signals to processor 1401 for processing. In this case, display screen 1405 may also be used to provide virtual buttons and / or virtual keyboards, also known as soft buttons and / or soft keyboards. In some embodiments, there may be one display screen 1405, disposed on the front panel of terminal 1400; in other embodiments, there may be at least two display screens 1405, disposed on different surfaces of terminal 1400 or in a foldable design; in other embodiments, display screen 1405 may be a flexible display, disposed on a curved surface or a foldable surface of terminal 1400. Display screen 1405 may even be configured as a non-rectangular irregular shape, i.e., a special-shaped screen. The display screen 1405 can be made of materials such as LCD (Liquid Crystal Display) and OLED (Organic Light-Emitting Diode).

[0550] The camera assembly 1406 is used to capture images or videos. Optionally, the camera assembly 1406 includes a front camera and a rear camera. Typically, the front camera is located on the front panel of the terminal, and the rear camera is located on the back of the terminal. In some embodiments, there are at least two rear cameras, each of which is a main camera, a depth-of-field camera, a wide-angle camera, and a telephoto camera, so as to achieve a background blur function by fusing the main camera and the depth-of-field camera, a panoramic shooting function by fusing the main camera and the wide-angle camera, and a VR (Virtual Reality) shooting function or other fusion shooting functions. In some embodiments, the camera assembly 1406 may also include a flash. The flash can be a single-color temperature flash or a dual-color temperature flash. A dual-color temperature flash refers to a combination of a warm light flash and a cold light flash, which can be used for light compensation at different color temperatures.

[0551] Audio circuit 1407 may include a microphone and a speaker. The microphone is used to collect sound waves from the user and the environment, converting them into electrical signals that are then input into processor 1401 for processing, or into RF circuit 1404 for voice communication. For stereo sound collection or noise reduction, multiple microphones may be provided, located in different locations within terminal 1400. The microphone may also be an array microphone or an omnidirectional microphone. The speaker is used to convert electrical signals from processor 1401 or RF circuit 1404 into sound waves. The speaker may be a traditional thin-film speaker or a piezoelectric ceramic speaker. A piezoelectric ceramic speaker can convert electrical signals not only into sound waves audible to humans, but also into sound waves inaudible to humans for purposes such as distance measurement. In some embodiments, audio circuit 1407 may also include a headphone jack.

[0552] Positioning component 1408 is used to locate the current geographic location of terminal 1400 to implement navigation or LBS (Location Based Service). Positioning component 1408 can be a positioning component based on the US GPS (Global Positioning System), China's Beidou system, or Russia's Galileo system.

[0553] Power supply 1409 is used to power various components in terminal 1400. Power supply 1409 can be AC ​​power, DC power, a disposable battery, or a rechargeable battery. When power supply 1409 includes a rechargeable battery, the rechargeable battery can be a wired rechargeable battery or a wireless rechargeable battery. A wired rechargeable battery is charged via a wired line, while a wireless rechargeable battery is charged via a wireless coil. The rechargeable battery can also support fast charging technology.

[0554] In some embodiments, the terminal 1400 further includes one or more sensors 1410 , including but not limited to: an acceleration sensor 1411 , a gyroscope sensor 1412 , a pressure sensor 1413 , a fingerprint sensor 1414 , an optical sensor 1415 , and a proximity sensor 1416 .

[0555] Accelerometer 1411 can detect the magnitude of acceleration along the three coordinate axes of the coordinate system established by terminal 1400. For example, accelerometer 1411 can be used to detect the components of gravity acceleration along the three coordinate axes. Processor 1401 can control display screen 1405 to display the user interface in either a landscape or portrait view based on the gravity acceleration signal collected by accelerometer 1411. Accelerometer 1411 can also be used to collect game or user motion data.

[0556] The gyroscope sensor 1412 can detect the orientation and rotation angle of the terminal 1400. It can work in conjunction with the accelerometer 1411 to collect the user's 3D movements on the terminal 1400. Based on the data collected by the gyroscope sensor 1412, the processor 1401 can implement the following functions: motion sensing (for example, changing the UI based on the user's tilt operation), image stabilization during shooting, game control, and inertial navigation.

[0557] The pressure sensor 1413 can be provided on the side frame of the terminal 1400 and / or below the display screen 1405. When the pressure sensor 1413 is provided on the side frame of the terminal 1400, it can detect the user's gripping signal of the terminal 1400, and the processor 1401 can perform left-hand or right-hand recognition or shortcut operations based on the gripping signal collected by the pressure sensor 1413. When the pressure sensor 1413 is provided below the display screen 1405, the processor 1401 controls the operable controls on the UI interface based on the user's pressure operation on the display screen 1405. Operable controls include at least one of a button control, a scroll bar control, an icon control, and a menu control.

[0558] Fingerprint sensor 1414 is used to collect the user's fingerprint. Processor 1401 identifies the user based on the fingerprint collected by fingerprint sensor 1414, or alternatively, fingerprint sensor 1414 identifies the user based on the collected fingerprint. Upon determining that the user's identity is trusted, processor 1401 authorizes the user to perform relevant sensitive operations, including unlocking the screen, viewing encrypted information, downloading software, making payments, and changing settings. Fingerprint sensor 1414 can be located on the front, back, or side of terminal 1400. If physical buttons or manufacturer logos are provided on terminal 1400, fingerprint sensor 1414 can be integrated with the physical buttons or manufacturer logos.

[0559] Optical sensor 1415 is used to detect ambient light intensity. In one embodiment, processor 1401 can control the display brightness of display screen 1405 based on the ambient light intensity detected by optical sensor 1415. Specifically, when the ambient light intensity is high, the display brightness of display screen 1405 is increased; when the ambient light intensity is low, the display brightness of display screen 1405 is decreased. In another embodiment, processor 1401 can also dynamically adjust the shooting parameters of camera assembly 1406 based on the ambient light intensity detected by optical sensor 1415.

[0560] Proximity sensor 1416, also known as a distance sensor, is typically located on the front panel of terminal 1400. Proximity sensor 1416 is used to detect the distance between the user and the front of terminal 1400. In one embodiment, when proximity sensor 1416 detects that the distance between the user and the front of terminal 1400 is gradually decreasing, processor 1401 controls display screen 1405 to switch from the screen-on state to the screen-off state. When proximity sensor 1416 detects that the distance between the user and the front of terminal 1400 is gradually increasing, processor 1401 controls display screen 1405 to switch from the screen-off state to the screen-on state.

[0561] Those skilled in the art will understand that Figure 14 The structure shown in the figure does not constitute a limitation on the terminal 1400, and the terminal 1400 may include more or fewer components than shown in the figure, or combine certain components, or adopt a different component arrangement.

[0562] The electronic device in the above method embodiment can be implemented as a server. For example, Figure 15 This is a schematic diagram of the structure of a server according to an exemplary embodiment. The server 1500 may have relatively large differences due to different configurations or performances. It can include one or more processors (Central Processing Units, CPU) 1501 and one or more memories 1502, wherein the memory 1502 stores at least one instruction, and the at least one instruction is loaded and executed by the processor 1501 to implement the multimedia resource recommendation method provided by the above-mentioned various method embodiments. Of course, the server can also have components such as wired or wireless network interfaces and input and output interfaces for input and output. The server can also include other components for implementing device functions, which will not be described here.

[0563] In an exemplary embodiment, a computer-readable storage medium including instructions is also provided, such as a memory including at least one instruction. The at least one instruction is executable by a processor to implement the multimedia resource recommendation method of the above embodiment. Alternatively, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc (CD-ROM), a magnetic tape, a floppy disk, or an optical data storage device.

[0564] In an exemplary embodiment, a computer program product is also provided, comprising one or more computer instructions stored in a computer-readable storage medium. One or more processors of an electronic device can read the one or more computer instructions from the computer-readable storage medium and execute the one or more computer instructions, thereby enabling the electronic device to perform the multimedia resource recommendation method described above.

[0565] In some embodiments, the instructions involved in the embodiments of the present disclosure may be deployed for execution on one electronic device, or on multiple electronic devices located at one location, or on multiple electronic devices distributed at multiple locations and interconnected through a communication network. Multiple electronic devices distributed at multiple locations and interconnected through a communication network may constitute a blockchain system.

[0566] Other embodiments of the present disclosure will readily occur to those skilled in the art after considering the specification and practicing the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, with the true scope and spirit of the present disclosure being indicated by the following claims.

[0567] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes can be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A multimedia resource recommendation method, characterized in that: include: Acquire a multimedia resource set that the user account has interacted with, the multimedia resource set including a plurality of first multimedia resources; For each candidate multimedia resource among a plurality of candidate multimedia resources, obtaining candidate association information between the first multimedia resource and the second multimedia resource based on a first resource category parameter of each first multimedia resource and a first similarity between a second multimedia resource of the first multimedia resource and the candidate multimedia resource, wherein the first resource category parameter is used to indicate a likelihood that the first multimedia resource can represent a corresponding category, and the second multimedia resource is a multimedia resource in the multimedia resource set whose similarity to the first multimedia resource satisfies a first similarity condition; determining, based on the candidate association information between the first multimedia resource and the second multimedia resource, a second resource category parameter of the first multimedia resource and a second similarity between the first multimedia resource and the candidate multimedia resource, wherein the second resource category parameter is a resource category parameter obtained by updating the first resource category parameter; determining association information between the first multimedia resource and the second multimedia resource based on the second resource category parameter of the first multimedia resource, the second similarity, the first target value, and candidate association information between the first multimedia resource and the second multimedia resource; Determining, from the multimedia resource set, based on the association information between the first multimedia resource and the second multimedia resource, a plurality of third multimedia resources corresponding to the candidate multimedia resource, wherein the association information corresponding to the plurality of third multimedia resources is not zero, a similarity between the third multimedia resource and the candidate multimedia resource satisfies a second similarity condition, and a first resource category parameter of the third multimedia resource satisfies a parameter condition; Based on the multiple third multimedia resources, multimedia resources are recommended to the user account.

2. The multimedia resource recommendation method according to claim 1, characterized in that: The determining, based on the candidate association information between the first multimedia resource and the second multimedia resource, a second resource category parameter of the first multimedia resource and a second similarity between the first multimedia resource and the candidate multimedia resource includes: Based on the candidate association information between the first multimedia resource and the second multimedia resource, updating the representation information of the first multimedia resource to obtain target representation information of the first multimedia resource; Based on the target representation information of the first multimedia resource, a first resource category parameter of the first multimedia resource is updated to obtain a second resource category parameter of the first multimedia resource; Based on the target representation information of the first multimedia resource, a second similarity between the first multimedia resource and the candidate multimedia resource is determined.

3. The multimedia resource recommendation method according to claim 1, characterized in that: The determining, based on the second resource category parameter of the first multimedia resource, the second similarity, the first target value, and the candidate association information between the first multimedia resource and the second multimedia resource, includes: determining importance indication information of the first multimedia resource according to the second resource category parameter of the first multimedia resource and the second similarity; In response to the importance indication information of the first multimedia resource being greater than or equal to a first target value, using the candidate association information between the first multimedia resource and the second multimedia resource as the association information between the first multimedia resource and the second multimedia resource; In response to the importance indication information of the first multimedia resource being less than a first target value, the association information between the first multimedia resource and the second multimedia resource is set to zero.

4. The multimedia resource recommendation method according to claim 1, characterized in that: The first target value is a value greater than the second target value in the importance indication information of each first multimedia resource, and the second target value is a minimum value of a target length and the number of the first multimedia resources.

5. The multimedia resource recommendation method according to any one of claims 1 to 4, characterized in that: The method for obtaining the first resource category parameter of the first multimedia resource includes: Determining a target category corresponding to the first multimedia resource; Acquire average representation information of the first multimedia resources of the target category according to the representation information of each first multimedia resource belonging to the target category; A third similarity between the representation information of the first multimedia resource and the average representation information is obtained, and the third similarity is used as a first resource category parameter of the first multimedia resource.

6. The multimedia resource recommendation method according to any one of claims 1 to 4, characterized in that: The recommending multimedia resources to the user account based on the plurality of third multimedia resources includes: Determining a target multimedia resource from the plurality of candidate multimedia resources based on the plurality of third multimedia resources and an interaction time sequence of the plurality of third multimedia resources; The target multimedia resource is recommended to the user account.

7. The multimedia resource recommendation method according to claim 6, characterized in that: The determining of a target multimedia resource from the plurality of candidate multimedia resources based on the plurality of third multimedia resources and the interaction time sequence of the plurality of third multimedia resources includes: Acquire an interaction sequence for each candidate multimedia resource, wherein the interaction sequence is composed of the plurality of third multimedia resources arranged in interaction time sequence; Obtaining a matching degree between the interaction sequence of each candidate multimedia resource and each candidate multimedia resource; The candidate multimedia resource with the greatest matching degree is used as the target multimedia resource.

8. The multimedia resource recommendation method according to claim 7, characterized in that: The obtaining of a matching degree between the interaction sequence of each candidate multimedia resource and each candidate multimedia resource includes: For each candidate multimedia resource, obtaining first representation information corresponding to the plurality of third multimedia resources according to the plurality of third multimedia resources, where the first representation information is used to represent a set consisting of the plurality of third multimedia resources; acquiring second representation information according to the importance indication information of each third multimedia resource and a change in the third multimedia resource in the interaction sequence, where the second representation information conforms to the change; A matching degree between the interaction sequence of the candidate multimedia resource and the candidate multimedia resource is acquired according to the first representation information, the second representation information, and the representation information of the candidate multimedia resource.

9. The multimedia resource recommendation method according to claim 8, characterized in that: The acquiring, according to the first representation information, the second representation information, and the representation information of the candidate multimedia resource, a matching degree between the interaction sequence of the candidate multimedia resource and the candidate multimedia resource includes: Obtaining a Hadamard product of the first representation information and the representation information of the candidate multimedia resource; Connecting the first representation information, the second representation information, the representation information of the candidate multimedia resource, and the Hadamard product to obtain third representation information; A dimensionality reduction process is performed on the third representation information to obtain a matching degree between the interaction sequence of the candidate multimedia resource and the candidate multimedia resource.

10. A multimedia resource recommendation device, characterized in that: include: An acquiring unit is configured to acquire a multimedia resource set that a user account has interacted with, wherein the multimedia resource set includes a plurality of first multimedia resources; a determining unit configured to, for each candidate multimedia resource among a plurality of candidate multimedia resources, obtain candidate association information between the first multimedia resource and the second multimedia resource based on a first resource category parameter of each first multimedia resource and a first similarity between a second multimedia resource of the first multimedia resource and the candidate multimedia resource, wherein the first resource category parameter is used to indicate a possibility that the first multimedia resource can represent a corresponding category, and the second multimedia resource is a multimedia resource in the multimedia resource set whose similarity to the first multimedia resource satisfies a first similarity condition; The determining unit is further configured to determine, based on the candidate association information between the first multimedia resource and the second multimedia resource, a second resource category parameter of the first multimedia resource and a second similarity between the first multimedia resource and the candidate multimedia resource, wherein the second resource category parameter is a resource category parameter obtained by updating the first resource category parameter; The determining unit is further configured to determine association information between the first multimedia resource and the second multimedia resource based on the second resource category parameter of the first multimedia resource, the second similarity, the first target value, and candidate association information between the first multimedia resource and the second multimedia resource; The determining unit is further configured to determine, from the multimedia resource set, a plurality of third multimedia resources corresponding to the candidate multimedia resource based on the association information between the first multimedia resource and the second multimedia resource, where the association information corresponding to the plurality of third multimedia resources is not zero, a similarity between the third multimedia resource and the candidate multimedia resource satisfies a second similarity condition, and a first resource category parameter of the third multimedia resource satisfies a parameter condition; The recommendation unit is configured to perform multimedia resource recommendation to the user account based on the multiple third multimedia resources.

11. The multimedia resource recommendation device according to claim 10, characterized in that: The determining unit is configured to perform: Based on the candidate association information between the first multimedia resource and the second multimedia resource, updating the representation information of the first multimedia resource to obtain target representation information of the first multimedia resource; Based on the target representation information of the first multimedia resource, a first resource category parameter of the first multimedia resource is updated to obtain a second resource category parameter of the first multimedia resource; Based on the target representation information of the first multimedia resource, a second similarity between the first multimedia resource and the candidate multimedia resource is determined.

12. The multimedia resource recommendation device according to claim 10, characterized in that: The determining unit is configured to perform: determining importance indication information of the first multimedia resource according to the second resource category parameter of the first multimedia resource and the second similarity; In response to the importance indication information of the first multimedia resource being greater than or equal to a first target value, using the candidate association information between the first multimedia resource and the second multimedia resource as the association information between the first multimedia resource and the second multimedia resource; In response to the importance indication information of the first multimedia resource being less than a first target value, the association information between the first multimedia resource and the second multimedia resource is set to zero.

13. The multimedia resource recommendation device according to claim 10, characterized in that: The first target value is a value greater than the second target value in the importance indication information of each first multimedia resource, and the second target value is a minimum value of a target length and the number of first multimedia resources in the multimedia resource set.

14. The multimedia resource recommendation device according to any one of claims 10 to 13, characterized in that: The method for obtaining the first resource category parameter of the first multimedia resource includes: Determining a target category corresponding to the first multimedia resource; Acquire average representation information of the first multimedia resources of the target category according to the representation information of each first multimedia resource belonging to the target category; A third similarity between the representation information of the first multimedia resource and the average representation information is obtained, and the third similarity is used as a first resource category parameter of the first multimedia resource.

15. The multimedia resource recommendation device according to any one of claims 10 to 13, characterized in that: The recommendation unit includes a determination module and a recommendation module; The determining module is configured to determine a target multimedia resource from the plurality of candidate multimedia resources based on the plurality of third multimedia resources and an interaction time sequence of the plurality of third multimedia resources; The recommendation module is configured to recommend the target multimedia resource to the user account.

16. The multimedia resource recommendation device according to claim 15, characterized in that: The determination module includes an acquisition submodule and a determination submodule; The acquisition submodule is configured to execute an interaction sequence for acquiring each candidate multimedia resource, wherein the interaction sequence is composed of the plurality of third multimedia resources arranged in interaction time sequence; The acquisition submodule is further configured to execute acquisition of a matching degree between the interaction sequence of each candidate multimedia resource and each candidate multimedia resource; The determination submodule is configured to execute the process of selecting the candidate multimedia resource with the greatest matching degree as the target multimedia resource.

17. The multimedia resource recommendation device according to claim 16, characterized in that: The acquisition submodule is configured to perform: For each candidate multimedia resource, obtaining first representation information corresponding to the plurality of third multimedia resources according to the plurality of third multimedia resources, where the first representation information is used to represent a set consisting of the plurality of third multimedia resources; acquiring second representation information according to the importance indication information of each third multimedia resource and a change in the third multimedia resource in the interaction sequence, where the second representation information conforms to the change; A matching degree between the interaction sequence of the candidate multimedia resource and the candidate multimedia resource is acquired according to the first representation information, the second representation information, and the representation information of the candidate multimedia resource.

18. The multimedia resource recommendation device according to claim 17, characterized in that: The acquisition submodule is configured to perform: Obtaining a Hadamard product of the first representation information and the representation information of the candidate multimedia resource; Connecting the first representation information, the second representation information, the representation information of the candidate multimedia resource, and the Hadamard product to obtain third representation information; A dimensionality reduction process is performed on the third representation information to obtain a matching degree between the interaction sequence of the candidate multimedia resource and the candidate multimedia resource.

19. An electronic device, characterized in that: include: one or more processors; one or more memories for storing the one or more processor-executable instructions; The processor is configured to execute the instructions to implement the multimedia resource recommendation method according to any one of claims 1 to 9. 20 . A computer-readable storage medium, wherein when instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the multimedia resource recommendation method according to claim 1 .

21. A computer program product comprising computer instructions, characterized in that When the computer instructions are executed by a processor, the multimedia resource recommendation method according to any one of claims 1 to 9 is implemented.

Images