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

Abstract

The present application relates to the field of computer technology and discloses a multimedia data recommendation method, device, electronic device and storage medium. The method is as follows: in response to an operation instruction initiated on an operable page for target multimedia data, a target association information feature vector corresponding to the target multimedia data is obtained. The target association information feature vector is constructed using a classification model based on the following information: the ID feature of the target multimedia data and the corresponding first weight, as well as the label feature group of the target multimedia data and the corresponding second weight. Then, a similar multimedia data set corresponding to the target multimedia data is obtained, and similar multimedia data is selected and presented on the operable page. In this way, the training difficulty of the classification model is reduced, the training speed of the classification model is increased, the construction speed of the association information feature vector corresponding to the multimedia data is accelerated, and the recommendation efficiency of the multimedia data is improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology and discloses a multimedia data recommendation method, device, electronic device and storage medium. Background Art

[0002] With the development of data processing technology, in order to explore the potential needs of target objects, it is usually necessary to recommend data that may be of interest to the target objects based on the target objects' operation behaviors on data in various scenarios.

[0003] In related technologies, when performing data recommendation, a model is usually used to construct a vector form corresponding to the data based on a fixed number of data features, and then determine the recommended data based on the similarity between the vectors.

[0004] However, when a fixed number of data features are input into the model for vector construction, the model needs to train the corresponding weight parameters for each data feature separately. Moreover, given that more data features can better represent the characteristics of the data, the number of input data features is usually set to a large value, which means that the model needs to learn too many weight parameters, greatly reducing the speed of vector generation and affecting the data recommendation effect. Summary of the Invention

[0005] The embodiments of the present application provide a multimedia data recommendation method, device, electronic device and storage medium to solve the problem that the model needs to learn too many weight parameters when constructing a vector, which affects the recommendation effect of the data.

[0006] In a first aspect, an embodiment of the present application provides a multimedia data recommendation method, the method comprising:

[0007] In response to an operation instruction initiated on an operable page for target multimedia data, obtaining a target-related information feature vector corresponding to the target multimedia data, the target-related information feature vector being constructed using a classification model based on the following information: an identity ID feature of the target multimedia data and a corresponding first weight, and a tag feature group of the target multimedia data and a corresponding second weight, wherein the first weight and the second weight are iteratively adjusted during a training process of the classification model;

[0008] Obtaining a similar multimedia data set established corresponding to the target multimedia data, and screening out first similar multimedia data that meets a second set condition from the similar multimedia data set, wherein the similar multimedia data set includes similar multimedia data, and similarity between a correlation information feature vector corresponding to the similar multimedia data and the target correlation information feature vector meets the first set condition;

[0009] The first similar multimedia data meeting the second set condition is used as multimedia data to be recommended and presented on the operable page.

[0010] Optionally, the classification model is trained in the following manner, including:

[0011] Based on first historical behavior data of each object within a first specified time period, a graph structure consisting of nodes and directed edges corresponding to the multimedia data is constructed, and an edge weight of each directed edge is determined. Furthermore, a random walk method is used to sample nodes included in the graph structure based on the edge weights of each directed edge to generate corresponding sequences of multimedia data samples, wherein the edge weight corresponding to the directed edge is positively correlated with the number of times two multimedia data corresponding to the directed edge are operated on in order of their directions.

[0012] Based on the multimedia data sample sequences, the classification model to be trained is subjected to multiple rounds of iterative training until a preset convergence condition is met; wherein, during one training process, the following operations are performed:

[0013] Determining a multimedia data input sample and a multimedia data output sample based on a currently obtained multimedia data sample sequence, and inputting an ID feature and a label feature group of the multimedia data input sample into a classification model to be trained, obtaining an associated information feature vector constructed corresponding to the multimedia data input sample, and obtaining a predicted classification result output by the classification model to be trained;

[0014] Based on the difference between the predicted classification result and the multimedia data output sample, a loss value is calculated, and a first weight corresponding to the ID feature setting and a second weight corresponding to the label feature group setting are adjusted by using gradient back propagation.

[0015] Optionally, inputting the ID feature and label feature group in the multimedia data input sample into the classification model to be trained to obtain the associated information feature vector constructed corresponding to the multimedia data input sample includes:

[0016] Inputting the ID features and the label feature group in the multimedia data input sample into a classification model to be trained, mapping the ID features into a first intermediate vector of a specified dimension through the classification model to be trained, and mapping the label feature group into a second intermediate vector of a specified dimension through the classification model to be trained;

[0017] The first intermediate vector and the second intermediate vector are fused based on the first weight corresponding to the ID feature and the second weight corresponding to the label feature group by the classification model to be trained to obtain an associated information feature vector constructed corresponding to the multimedia data input sample.

[0018] Optionally, mapping the label feature group into a second intermediate vector of a specified dimension by using the classification model to be trained includes:

[0019] Splitting the label feature group based on a preset separator using the classification model to be trained to obtain individual label features, and mapping each label feature into a second intermediate sub-vector of a corresponding specified dimension;

[0020] The elements at each position in each second intermediate sub-vector are determined respectively, and the average value of each element at the same position in each second intermediate sub-vector is used as the first new element at the corresponding position, and based on the first new elements corresponding to each position, a second intermediate vector of the specified dimension corresponding to the label group is generated.

[0021] Optionally, also include:

[0022] When it is determined that there is a newly added label feature in the label feature group, the label feature group is split based on a preset separator by the classification model to be trained to obtain each label feature and each newly added label feature;

[0023] Mapping each newly added label feature to a second intermediate newly added sub-vector of a specified dimension respectively corresponding to each label feature through the classification model to be trained, and determining a second intermediate sub-vector of the specified dimension respectively corresponding to each label feature;

[0024] Determine each second intermediate newly added sub-vector and each element at each position in each second intermediate sub-vector, and use the average value of each element at the same position in each second intermediate newly added sub-vector and each second intermediate sub-vector as the second new element at the corresponding position, and generate a second intermediate vector of the specified dimension corresponding to the label feature group based on the second new elements corresponding to each position.

[0025] Optionally, constructing a graph structure consisting of nodes and directed edges corresponding to multimedia data based on the first historical behavior data of each object within the first specified time period includes:

[0026] Obtaining first historical behavior data of each object within a first specified time period, and extracting a historical multimedia data sequence of each object's respective operations within the first specified time period from the first historical behavior data, wherein each multimedia data in the historical multimedia data sequence is arranged in the order in which it was operated, and each multimedia data is associated with time information of the operation;

[0027] Splitting the historical multimedia data sequences corresponding to the respective objects according to the specified time lengths to obtain at least one multimedia data subsequence associated with the respective objects;

[0028] Generate corresponding nodes according to the multimedia data included in each obtained multimedia data subsequence, and connect the corresponding nodes according to the order in which the multimedia data are operated to obtain corresponding directed edges, wherein one multimedia data corresponds to one node in the graph structure;

[0029] Based on the nodes and the directed edges obtained, the construction of the graph structure is completed.

[0030] Optionally, the random walk method is used to sample each node included in the graph structure based on the edge weight of each directed edge to generate corresponding multimedia data sample sequences, including:

[0031] For each directed edge in the graph structure, performing the following operations: determining, based on the direction of a directed edge, a sequence in which two corresponding multimedia data items are operated, and using a number of operations in the first historical behavior data that matches the sequence in which the two multimedia data items are operated as an edge weight of the corresponding directed edge;

[0032] A random walk algorithm from node to vector node2vec is adopted, and based on the edge weights corresponding to the directed edges, sampling is performed in each node of the graph structure to generate corresponding multimedia data sample sequences.

[0033] Optionally, determining the multimedia data input sample and the multimedia data output sample based on the multimedia data sample sequence includes:

[0034] For each multimedia data sample sequence, perform the following operations:

[0035] Sequentially acquiring, according to the arrangement order of the multimedia data included in a multimedia data sample sequence, the multimedia data except the last multimedia data in the multimedia data sample sequence;

[0036] The ID feature and label feature group corresponding to any acquired multimedia data is used as a multimedia data input sample;

[0037] Another multimedia data that is subsequently operated after the any multimedia data is acquired, and a classification result corresponding to the another multimedia data is used as a multimedia data output sample corresponding to the multimedia data input sample.

[0038] Optionally, obtaining a target association information feature vector corresponding to the target multimedia data includes:

[0039] obtaining a target association information feature vector constructed by the classification model to be trained when the classification model to be trained is trained using the target multimedia data; or

[0040] A target association information feature vector constructed based on the ID feature and label feature group of the target multimedia data using a trained classification model is obtained.

[0041] Optionally, before responding to the operation instruction initiated on the operable page for the target multimedia data, the method further includes:

[0042] Obtaining a target-related information feature vector corresponding to the target multimedia data, and based on similarities between the target-related information feature vector and other related information feature vectors corresponding to other multimedia data, screening out N multimedia data with the highest similarity to the target-related information feature vector as similar multimedia data that meet a first set condition, and generating a similar multimedia data set corresponding to the target multimedia data based on each similar multimedia data;

[0043] The step of screening out first similar multimedia data that meets a second set condition from the similar multimedia data set includes:

[0044] In descending order of similarity with the target association information feature vector, M similar multimedia data are obtained from the similar multimedia data set as first similar multimedia data that meet the second set condition, where M is a preset integer.

[0045] Optionally, also include:

[0046] In response to a refresh operation on the operable page, filtering out multimedia data of interest to the object based on second historical behavior data of the object that logged into the operable page within a second specified time period, and obtaining a set of similar multimedia data corresponding to the multimedia data of interest;

[0047] From a set of similar multimedia data corresponding to the multimedia data of interest, a specified number of second similar multimedia data are screened out and presented on the operable page, and in response to an operation instruction initiated on the operable page for the target multimedia data, a step of obtaining a target association information feature vector corresponding to the target multimedia data is performed.

[0048] In a second aspect, an embodiment of the present application provides a multimedia data recommendation device, comprising:

[0049] a response unit, configured to, in response to an operation instruction initiated on an operable page for target multimedia data, obtain a target association information feature vector corresponding to the target multimedia data, wherein the target association information feature vector is constructed using a classification model based on the following information: an identity ID feature of the target multimedia data and a corresponding first weight, and a tag feature group of the target multimedia data and a corresponding second weight, wherein the first weight and the second weight are iteratively adjusted during a training process of the classification model;

[0050] an acquiring unit, configured to acquire a similar multimedia data set established corresponding to the target multimedia data, and screen out first similar multimedia data satisfying a second set condition from the similar multimedia data set, wherein the similar multimedia data set includes similar multimedia data, and a similarity between a correlation information feature vector corresponding to the similar multimedia data and the target correlation information feature vector satisfies the first set condition;

[0051] The presenting unit is configured to present the first similar multimedia data satisfying the second set condition as multimedia data to be recommended on the operable page.

[0052] Optionally, the device further includes a training unit for training the classification model in the following manner:

[0053] Based on first historical behavior data of each object within a first specified time period, a graph structure consisting of nodes and directed edges corresponding to the multimedia data is constructed, and an edge weight of each directed edge is determined. Furthermore, a random walk method is used to sample nodes included in the graph structure based on the edge weights of each directed edge to generate corresponding sequences of multimedia data samples, wherein the edge weight corresponding to the directed edge is positively correlated with the number of times two multimedia data corresponding to the directed edge are operated on in order of their directions.

[0054] Based on the multimedia data sample sequences, the classification model to be trained is subjected to multiple rounds of iterative training until a preset convergence condition is met; wherein, during one training process, the following operations are performed:

[0055] Determining a multimedia data input sample and a multimedia data output sample based on a currently obtained multimedia data sample sequence, and inputting an ID feature and a label feature group of the multimedia data input sample into a classification model to be trained, obtaining an associated information feature vector constructed corresponding to the multimedia data input sample, and obtaining a predicted classification result output by the classification model to be trained;

[0056] Based on the difference between the predicted classification result and the multimedia data output sample, a loss value is calculated, and a first weight corresponding to the ID feature setting and a second weight corresponding to the label feature group setting are adjusted by using gradient back propagation.

[0057] Optionally, when inputting the ID feature and label feature group in the multimedia data input sample into the classification model to be trained to obtain the association information feature vector constructed corresponding to the multimedia data input sample, the training unit is used to:

[0058] Inputting the ID features and the label feature group in the multimedia data input sample into a classification model to be trained, mapping the ID features into a first intermediate vector of a specified dimension through the classification model to be trained, and mapping the label feature group into a second intermediate vector of a specified dimension through the classification model to be trained;

[0059] The first intermediate vector and the second intermediate vector are fused based on the first weight corresponding to the ID feature and the second weight corresponding to the label feature group by the classification model to be trained to obtain an associated information feature vector constructed corresponding to the multimedia data input sample.

[0060] Optionally, when mapping the label feature group into a second intermediate vector of a specified dimension through the classification model to be trained, the training unit is configured to:

[0061] Splitting the label feature group based on a preset separator using the classification model to be trained to obtain individual label features, and mapping each label feature into a second intermediate sub-vector of a corresponding specified dimension;

[0062] The elements at each position in each second intermediate sub-vector are determined respectively, and the average value of each element at the same position in each second intermediate sub-vector is used as the first new element at the corresponding position, and based on the first new elements corresponding to each position, a second intermediate vector of the specified dimension corresponding to the label group is generated.

[0063] Optionally, the training unit is further configured to:

[0064] When it is determined that there is a newly added label feature in the label feature group, the label feature group is split based on a preset separator by the classification model to be trained to obtain each label feature and each newly added label feature;

[0065] Mapping each newly added label feature to a second intermediate newly added sub-vector of a specified dimension respectively corresponding to each label feature through the classification model to be trained, and determining a second intermediate sub-vector of the specified dimension respectively corresponding to each label feature;

[0066] Determine each second intermediate newly added sub-vector and each element at each position in each second intermediate sub-vector, and use the average value of each element at the same position in each second intermediate newly added sub-vector and each second intermediate sub-vector as the second new element at the corresponding position, and generate a second intermediate vector of the specified dimension corresponding to the label feature group based on the second new elements corresponding to each position.

[0067] Optionally, when constructing a graph structure consisting of nodes and directed edges corresponding to multimedia data based on the first historical behavior data of each object within the first specified time period, the training unit is configured to:

[0068] Obtaining first historical behavior data of each object within a first specified time period, and extracting a historical multimedia data sequence of each object's respective operations within the first specified time period from the first historical behavior data, wherein each multimedia data in the historical multimedia data sequence is arranged in the order in which it was operated, and each multimedia data is associated with time information of the operation;

[0069] Splitting the historical multimedia data sequences corresponding to the respective objects according to the specified time lengths to obtain at least one multimedia data subsequence associated with the respective objects;

[0070] Generate corresponding nodes according to the multimedia data included in each obtained multimedia data subsequence, and connect the corresponding nodes according to the order in which the multimedia data are operated to obtain corresponding directed edges, wherein one multimedia data corresponds to one node in the graph structure;

[0071] Based on the nodes and the directed edges obtained, the construction of the graph structure is completed.

[0072] Optionally, when sampling in each node included in the graph structure based on the edge weights of each directed edge in a random walk manner to generate corresponding multimedia data sample sequences, the training unit is configured to:

[0073] For each directed edge in the graph structure, performing the following operations: determining, based on the direction of a directed edge, a sequence in which two corresponding multimedia data items are operated, and using a number of operations in the first historical behavior data that matches the sequence in which the two multimedia data items are operated as an edge weight of the corresponding directed edge;

[0074] A random walk algorithm from node to vector node2vec is adopted, and based on the edge weights corresponding to the directed edges, sampling is performed in each node of the graph structure to generate corresponding multimedia data sample sequences.

[0075] Optionally, when determining the multimedia data input sample and the multimedia data output sample based on the multimedia data sample sequence, the training unit is configured to:

[0076] For each multimedia data sample sequence, perform the following operations:

[0077] Sequentially acquiring, according to the arrangement order of the multimedia data included in a multimedia data sample sequence, the multimedia data except the last multimedia data in the multimedia data sample sequence;

[0078] The ID feature and label feature group corresponding to any acquired multimedia data is used as a multimedia data input sample;

[0079] Another multimedia data that is subsequently operated after the any multimedia data is acquired, and a classification result corresponding to the another multimedia data is used as a multimedia data output sample corresponding to the multimedia data input sample.

[0080] Optionally, when obtaining the target association information feature vector corresponding to the target multimedia data, the response unit is configured to:

[0081] obtaining a target association information feature vector constructed by the classification model to be trained when the classification model to be trained is trained using the target multimedia data; or

[0082] A target association information feature vector constructed based on the ID feature and label feature group of the target multimedia data using a trained classification model is obtained.

[0083] Optionally, before responding to an operation instruction initiated on the operable page for the target multimedia data, the response unit is further configured to:

[0084] Obtaining a target-related information feature vector corresponding to the target multimedia data, and based on similarities between the target-related information feature vector and other related information feature vectors corresponding to other multimedia data, screening out N multimedia data with the highest similarity to the target-related information feature vector as similar multimedia data that meet a first set condition, and generating a similar multimedia data set corresponding to the target multimedia data based on each similar multimedia data;

[0085] When the first similar multimedia data meeting the second set condition is screened out from the similar multimedia data set, the acquiring unit is configured to:

[0086] In descending order of similarity with the target association information feature vector, M similar multimedia data are obtained from the similar multimedia data set as first similar multimedia data that meet the second set condition, where M is a preset integer.

[0087] Optionally, the response unit is further configured to:

[0088] In response to a refresh operation on the operable page, filtering out multimedia data of interest to the object based on second historical behavior data of the object that logged into the operable page within a second specified time period, and obtaining a set of similar multimedia data corresponding to the multimedia data of interest;

[0089] From a set of similar multimedia data corresponding to the multimedia data of interest, a specified number of second similar multimedia data are screened out and presented on the operable page, and in response to an operation instruction initiated on the operable page for the target multimedia data, a step of obtaining a target association information feature vector corresponding to the target multimedia data is performed.

[0090] In a third aspect, an embodiment of the present application provides an electronic device, which includes at least one processor and at least one memory, wherein the memory stores program code, and when the program code is executed by the processor, the at least one processor is able to execute the multimedia data recommendation method provided in the embodiment of the present application.

[0091] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, which includes a program code. When the program product is run on an electronic device, the program code is used to enable the electronic device to execute the multimedia data recommendation method provided by the embodiment of the present application.

[0092] The beneficial effects of this application are as follows:

[0093] The embodiments of the present application provide a multimedia data recommendation method, device, electronic device and storage medium. In response to an operation instruction initiated on an operable page for target multimedia data, a target association information feature vector corresponding to the target multimedia data is obtained, and the target association information feature vector is constructed using a classification model based on the following information: the ID feature of the target multimedia data and the corresponding first weight, and the label feature group of the target multimedia data and the corresponding second weight, wherein the first weight and the second weight are iteratively adjusted during the training process of the classification model, and then a similar multimedia data set corresponding to the target multimedia data is obtained, and first similar multimedia data that meets a second set condition is screened out from the similar multimedia data set, wherein the similar multimedia data set includes similar multimedia data, and the similarity between the association information feature vector corresponding to the similar multimedia data and the target association information feature vector meets the first set condition, and then the first similar multimedia data that meets the second set condition is presented as the multimedia data to be recommended on the operable page.

[0094] In this way, the features of the multimedia data input into the classification model are divided into two categories, namely ID features and label feature groups, and a first weight is set corresponding to the ID feature in the classification model, and a second weight is set corresponding to the label feature group. This avoids the problem of needing to learn the corresponding weights for each independent feature when each independent feature of the data is input into the model separately under related technologies. The classification model only needs to learn the first weight and the second weight during the training process, which reduces the training difficulty of the classification model, improves the training speed of the classification model, speeds up the construction speed of the associated information feature vector corresponding to the multimedia data, and improves the recommendation efficiency of multimedia data.

[0095] Other features and advantages of the present application will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present application. The purposes and other advantages of the present application can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0097] Figure 1a This is a schematic diagram of an application scenario in an embodiment of the present application;

[0098] Figure 1b A schematic diagram of an operable page for presenting multimedia data in an embodiment of the present application;

[0099] Figure 2a This is a schematic diagram of the architecture of the classification model in the embodiment of the present application;

[0100] Figure 2b Schematic diagram of the training process of the classification model in the embodiment of the present application;

[0101] Figure 2c A schematic diagram of a process for constructing a multimedia data sample sequence in an embodiment of the present application;

[0102] Figure 2d This is a schematic diagram of a historical multimedia data sequence in an embodiment of the present application;

[0103] Figure 2e This is a schematic diagram of splitting a historical multimedia data sequence into at least one multimedia data subsequence in an embodiment of the present application;

[0104] Figure 2f A schematic diagram of nodes corresponding to multimedia data generated in an embodiment of the present application;

[0105] Figure 2g This is a flow chart of a training process in an embodiment of the present application;

[0106] Figure 2h Schematic diagram of the process of generating the associated information feature vector in an embodiment of the present application;

[0107] Figure 3 This is a schematic diagram of the multimedia data recommendation process in an embodiment of the present application;

[0108] Figure 4a This is a schematic diagram of an operational page presented in an embodiment of the present application;

[0109] Figure 4b This is a schematic diagram of an operable page further presented in an embodiment of the present application;

[0110] Figure 4c A schematic diagram of a playback interface for multimedia data presented in an embodiment of the application;

[0111] Figure 4d This is a schematic diagram of the process of recommending electronic reading materials in an embodiment of the present application;

[0112] Figure 5 A schematic diagram of the logical structure of a multimedia data recommendation device provided in an embodiment of the present application;

[0113] Figure 6 A schematic diagram of the hardware structure of an electronic device to which an embodiment of the present application is applied;

[0114] Figure 7This is a schematic diagram of the structure of a computing device in an embodiment of the present application. DETAILED DESCRIPTION

[0115] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of the technical solutions of this application, but not all of them. Based on the embodiments described in this application document, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the technical solutions of this application.

[0116] The terms "first," "second," and the like in the specification and claims of this application and in the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a particular order or sequential sequence. It should be understood that the terms used in this manner are interchangeable where appropriate, such that the embodiments of the invention described herein can be practiced in sequences other than those illustrated or described herein.

[0117] The following explains some of the terms used in the embodiments of the present application to facilitate understanding by those skilled in the art.

[0118] Multimedia data, in the embodiment of this application, refers to data composed of at least one content such as video, audio, and text. In this application, multimedia data can be text data, such as e-books, news, papers, etc., or various types of audio data, such as music, crosstalk audio, etc.; it can also be video data, such as documentaries, crosstalk videos, news videos, etc.; it can also be data obtained by combining text data and audio data, such as audio books, etc.

[0119] Word embedding: Word embedding refers to a method of representing a data or object using a low-dimensional dense vector. In the embodiment of the present application, the distance between the constructed vectors can reflect the similarity between different multimedia data.

[0120] Node-to-vector algorithm: The node-to-vector algorithm (node2vec) can sample nodes in the graph structure in a random walk manner to obtain various node sequences. In the embodiment of the present application, each node in the graph structure corresponds to a multimedia data, and then corresponds to each node sequence, and various multimedia data sequences can be obtained.

[0121] The embodiments of the present application relate to artificial intelligence (AI) and machine learning technology, and are designed based on computer vision technology and machine learning (ML) in artificial intelligence.

[0122] Artificial intelligence (AI) refers to the theories, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. AI also involves studying the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making.

[0123] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, encompassing both hardware and software technologies. Foundational AI technologies generally include sensors, specialized AI chips, cloud computing, distributed storage, big data processing, interactive systems, and mechatronics. Software technologies typically encompass computer vision, natural language processing, and machine learning / deep learning. With the development and advancement of AI, research and application are being conducted in numerous fields, including smart homes, smart customer service, virtual assistants, smart speakers, smart marketing, driverless and autonomous driving, robotics, and smart healthcare. With further technological advancements, AI will be applied in even more areas, playing an increasingly important role.

[0124] Machine learning (ML) is a multidisciplinary field that encompasses probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computers can simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is at the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications span all areas of AI. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and self-learning.

[0125] The classification models in the multimedia data recommendation method proposed in the embodiments of the present application are all machine learning models. These models involve the technical field of machine learning, and the classification models can be trained through machine learning technology.

[0126] The following is a brief introduction to the design concept of the embodiment of this application:

[0127] Under related technologies, when recommending content that may be of interest to a target object, a data sample sequence is usually randomly constructed based on the historical operation data of each target object, and the data sample sequence is used to train a model. A vector corresponding to the data is constructed through the model, and by calculating the similarity between the vectors, data similar to the target data of the operation is determined, and the determined similar data is used as the recommended data. In this way, when constructing the vector corresponding to the data, a specified number of features corresponding to the data need to be input into the model, and a vector corresponding to the data is constructed based on the weights corresponding to each feature in the model. In this way, considering that the more features used to describe the data, the clearer the characteristics of a data can be represented, when more features are used to describe a data, the more weights the model needs to learn, which greatly affects the processing speed of the model and indirectly reduces the processing efficiency of data recommendation.

[0128] In view of this, in an embodiment of the present application, the processing device divides the features of the multimedia data input into the classification model into two categories, namely identity identification (ID) features and label feature groups, and sets a first weight corresponding to the ID features in the classification model, and sets a second weight corresponding to the label feature group. This avoids the problem of needing to learn the corresponding weights for each independent feature when inputting each independent feature of the data into the model separately under related technologies, so that the classification model only needs to learn the first weight and the second weight during the training process, which reduces the training difficulty of the classification model, improves the training speed of the classification model, accelerates the construction speed of the associated information feature vector corresponding to the multimedia data, and improves the recommendation efficiency of multimedia data.

[0129] The preferred embodiments of the present application are described below in conjunction with the drawings in the specification. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application and are not used to limit the present application. In addition, the embodiments of the present application and the features in the embodiments may be combined with each other if there is no conflict.

[0130] like Figure 1a , which is a schematic diagram of an application scenario in an embodiment of the present application. The application scenario diagram includes two terminal devices 110 and a server 120, 1101 is an operable page on the terminal device, and the terminal device 110 and the server 120 can communicate through a communication network.

[0131] In an optional embodiment, the communication network is a wired network or a wireless network. The terminal device 110 and the server 120 can be directly or indirectly connected via wired or wireless communication, which is not limited in this application.

[0132] In the embodiment of the present application, the terminal device 110 is an electronic device used by the target object, which can be a personal computer, mobile phone, tablet computer, notebook, desktop computer, smart watch, e-book reader, etc., a computer device with certain computing capabilities and running instant messaging software and websites or social software and websites, but is not limited to this.

[0133] Server 120 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It can also be 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, information recommendation networks (Content Delivery Network, CDN), as well as big data and artificial intelligence platforms.

[0134] The terminal device 110 and the server 120 are connected directly or indirectly using wired or wireless communication, which is not limited in this application. The terminal device 110 is installed with an application related to multimedia data recommendation, presenting an operational page 1101. The application involved in the embodiment of the present application can be software, or a client such as a web page or mini-program. The server 120 can be a backend server corresponding to the software, web page, mini-program, etc., or other server that can obtain backend data.

[0135] In an embodiment of the present application, when recommending multimedia data, the processing device may be a server 120. For example, when recommending multimedia data, the server 120, as a processing device, receives an operation instruction initiated by the terminal device 110 for the target multimedia data, thereby determining the recommended first similar multimedia data, and presenting the recommended first similar multimedia data in an operable interface on the terminal device 110. Alternatively, the functions of the processing device may be implemented by a combination of the terminal device 110 and the server 120, with the terminal device 110 participating in executing some operations in the multimedia data recommendation process. For example, after the server 120 determines the similarity between the target associated information feature vector and the other associated information feature vectors corresponding to each other multimedia data, the server 120 may send the calculated similarities to the terminal device 110, and the terminal device 110 completes the screening of the first similar multimedia data. Of course, these processes may also be performed by the terminal device 110 alone, and are not specifically limited here.

[0136] See Figure 1bAs shown, it is a schematic diagram of an operable page for presenting multimedia data in an embodiment of the present invention. The operable page 1101 on the terminal device 110 includes at least an area that can respond to input operations of an object, specifically a search box, and an area that can present recommended multimedia data. When the multimedia data is operated, it can trigger the presentation of a playback page for the corresponding multimedia data. When the target object triggers a selection operation for multimedia data on the operable page 1101, it is considered that an operation instruction for the multimedia data has been initiated, and then the processing device executes the recommended operation of the multimedia data proposed in the present invention and presents the determined multimedia data to be recommended on the operable page 1101.

[0137] In one possible application scenario, the multimedia data in this application can be stored using cloud storage technology. Cloud storage is a new concept that extends and develops from the concept of cloud computing. A distributed cloud storage system (hereinafter referred to as a storage system) refers to a storage system that uses cluster applications, grid technology, and distributed storage file systems to bring together a large number of different types of storage devices (also known as storage nodes) in a network through application software or application interfaces to work together and provide external data storage and service access functions.

[0138] The following describes the multimedia data recommendation process in the embodiment of the present application with reference to the accompanying drawings:

[0139] It should be noted that in the embodiments of the present application, each multimedia data item is pre-configured with an ID feature and a tag feature group. The ID feature is generated to distinguish different multimedia data items and is used to identify different multimedia data items. Based on the ID feature, a corresponding multimedia data item can be uniquely identified. The tag feature group includes various types of tags set for the multimedia data, such as classification tags constructed according to the operational classification requirements of the data, content tags and entity tags extracted from the text content associated with the multimedia data, and so on. The classification tags can be determined by manual or model annotation of the multimedia data, while the content tags and entity tags can be determined by using a text tag classification algorithm and learned through natural language processing methods.

[0140] In the embodiments of the present application, the text tag classification algorithm used can be a commonly used weighted technique for information retrieval data mining (term frequency–inverse document frequency, TF-IDF), or a keyword extraction and summarization algorithm (TextRank), etc. In a specific implementation, to ensure that the text tag classification algorithm is used to accurately extract the characteristics of each tag used to describe multimedia data, the text tag classification algorithm can be pre-trained using annotated text corpus.

[0141] In particular, in the implementation of this application, when a natural language processing algorithm is used to extract content tags and entity tags based on the text content corresponding to the multimedia data, when it is determined that the multimedia data is audio data or video data, speech recognition can be used to convert the audio content into corresponding text content, and the content tags and entity tags are determined from the converted text content and other related text content, wherein the other text content includes the title, comments and other data of the multimedia data. The content tags are used to represent the main information to be expressed by the multimedia data, and the entity tags are used to represent the object entities related to the multimedia data.

[0142] For example, taking the multimedia data of the crosstalk "The Third Master Teaching His Son" as an example, the ID feature configured for the multimedia data "The Third Master Teaching His Son" is "TRA_3110002026_572", and the extracted label feature group is: "crosstalk sketch; Sun Y; Feng Zyang; Third Master", among which the label feature group includes the classification feature "crosstalk sketch" and the entity features "Sun Y", "Feng Zxiang", and "Third Master", and different label features are separated by ";" in the label feature group.

[0143] The following first describes the training process of the classification model with reference to the accompanying drawings. During the training process and after the training is completed, the classification model can construct an associated information feature vector corresponding to the input multimedia data based on the ID features and label features of the input multimedia data.

[0144] It should be noted that, in the embodiment of the present application, the processing device constructs a classification model for realizing the data classification function in advance according to the actual classification processing needs.

[0145] See Figure 2a As shown, it is a schematic diagram of the architecture of the classification model in an embodiment of the present application. The classification model built by the processing device includes at least an input layer, an embedding layer, and an output layer. In the training process, the processing device receives the ID features and label feature groups of the multimedia data input sample through the input layer of the classification model, and maps the ID features of the multimedia data input sample to vectors of specified dimensions through the embedding layer of the classification model, and maps each label feature in the label feature group to a vector of specified dimensions, and then averages the vectors corresponding to each label feature to obtain a vector of specified dimensions corresponding to the label feature group, and then obtains the associated information feature vector corresponding to the multimedia data input sample based on the first weight and the second weight to be learned, and obtains the classification result for each preset category after the dimension conversion of the associated information feature vector obtained as the dimension based on the normalized exponential function algorithm (softmax).

[0146] It should be noted that in the embodiment of the present application, when fusing the vector of the specified dimension corresponding to the ID feature and the vector of the specified dimension corresponding to the label feature group, a linear weighted average fusion method is used to obtain the associated information feature vector of the specified dimension. However, in actual processing, it can be expanded to any network layer that supports gradient back propagation, such as a fully connected layer, to replace the linear weighted average operation.

[0147] It should be noted that training the classification model can be a periodic process. The processing device can periodically generate training samples based on the first historical behavior data of each object globally to train the classification model. Furthermore, the training time can be selectively selected during periods when most objects are inactive. This allows the classification model to reconstruct the associated information feature vectors for the multimedia data using the associated information feature vectors, thereby updating the existing associated information feature vectors for the multimedia data. This allows the recommendation process to be based on the latest associated information feature vectors for each multimedia data item.

[0148] For example, the processing device may obtain the first historical behavior data of each object generated within a first specified time period in the past every day, generate training samples, and train the classification model, wherein the first historical behavior data may come from the operation log of the corresponding object.

[0149] For example, assuming that the first specified time period is 30 days and the current date is June 14, the processing device generates training samples for training the classification model based on the first historical behavior data of each object between May 15 and June 13.

[0150] For another example, assuming that the current date is June 15, the processing device obtains the first historical behavior data of each object generated between May 16 and June 14, and generates training samples for training the classification model.

[0151] In the following description, the training process of the classification model within one training cycle will be taken as an example to explain the training process of the classification model in detail. For one training process, the processing device may process multiple multimedia data input samples at the same time. For the convenience of description, the following description will only take the learning and training process of the classification model based on one multimedia data input sample as an example for schematic description. For the case where multiple multimedia data input samples are input into the classification model at the same time, the same processing process is performed in parallel in the classification model, which will not be repeated here.

[0152] See Figure 2b As shown, it is a schematic diagram of the training process of the classification model in the embodiment of the present application. Figure 2b , the training process of the classification model is explained in detail:

[0153] Step 201: The processing device constructs a graph structure consisting of nodes and directed edges corresponding to multimedia data based on the first historical behavior data of each object within the first specified time period, determines the edge weight of each directed edge, and adopts a random walk method to sample each node included in the graph structure based on the edge weight of each directed edge to generate corresponding multimedia data sample sequences.

[0154] When generating various multimedia data sample sequences for training a classification model, the processing device first constructs a graph structure consisting of nodes corresponding to the multimedia data and directed edges consisting of the operation sequence between the multimedia data based on the first historical behavior data of each object obtained within the first specified time period, and then generates various multimedia data sample sequences by performing random walk sampling in the nodes in the graph structure.

[0155] See Figure 2c As shown, it is a flow chart of constructing a multimedia data sample sequence in an embodiment of the present application. Figure 2c , a vector description of the generation process of each multimedia data sample sequence is given.

[0156] Step 2011: The processing device obtains the first historical behavior data of each object within the first specified time period, and extracts the historical multimedia data sequence of each object's respective operation within the first specified time period from the first historical behavior data, wherein each multimedia data in the historical multimedia data sequence is arranged in the order of being operated, and each multimedia data is associated with the time information of being operated.

[0157] Specifically, the processing device may use the operation log of each object within the first specified time period as the first historical behavior data of each object within the first specified time period, and use data cleaning to remove operations recorded in the operation log that are not related to multimedia data, and obtain a historical multimedia data sequence generated based on the multimedia data operated by each object within the first specified time period, wherein each multimedia data in the historical multimedia sequence is arranged in the order of being operated, and each multimedia data is associated with the time information of being operated.

[0158] For example, see Figure 2d As shown, it is a schematic diagram of a historical multimedia data sequence in an embodiment of the present application. The processing device generates a historical multimedia data sequence of each object based on the first historical behavior data of each object in a first specified time period, such as Figure 2d As shown, Ui, Uj, and Uk represent different objects respectively, and the historical multimedia data sequence is composed of multimedia data arranged in the order of operation time. Figure 2dThe multimedia data corresponding to different operation time points are only schematically presented.

[0159] In this way, a historical multimedia data sequence can be established according to the order in which the multimedia data are operated. With the help of the historical multimedia data sequence, the order in which different multimedia data are operated can be represented, the relationship between different multimedia data can be preliminarily established, and a basis can be provided for the generation of subsequent multimedia data sample sequences.

[0160] Step 2012: The processing device splits the historical multimedia data sequence corresponding to each object according to the specified time length, and obtains at least one multimedia data subsequence associated with each object.

[0161] After obtaining the historical multimedia data sequence corresponding to each object, the processing device splits the historical multimedia data sequence corresponding to each object according to the specified time length to obtain at least one multimedia data subsequence corresponding to each object.

[0162] In a specific embodiment, the processing device can use a specified time length as a time window to truncate the historical multimedia data sequence of each object into at least one multimedia data subsequence, thereby obtaining a multimedia data subsequence that can reflect the operational interest of each object within the specified time length.

[0163] It should be noted that in a possible embodiment of the present application, the specified duration can be set as the size of the time window. When truncating historical multimedia data, the time window can be slid on the time axis corresponding to the historical multimedia data sequence according to a certain step size, and the various multimedia data within the time range corresponding to the sliding time window can be combined into a multimedia data subsequence.

[0164] For example, assuming that the specified duration is 2 hours and the step size is 0.5 hours, then the 2-hour time window is used to slide on the historical multimedia data sequence. Assuming that the time range corresponding to the current time window position is 10:00-12:00, then when moving according to the step size of 0.5 hours, the time range covered by the next time window is 10:30-12:30, and then the sequence composed of each multimedia data with an operation time between 10:30-12:30 is regarded as a multimedia data subsequence.

[0165] Optionally, when determining a multimedia data subsequence, the time may be directly divided according to a set duration to obtain various time periods, and the sequence consisting of the multimedia data included in each time period may be used as the multimedia data subsequence within the corresponding time period.

[0166] For example, participation Figure 2eAs shown in FIG. , which is a schematic diagram of splitting a historical multimedia data sequence into at least one multimedia data subsequence in an embodiment of the present application, assuming that the specified time length is set to 3 hours, the historical multimedia data sequence obtained for any object Um is: ... multimedia data C- multimedia data D- multimedia data A- multimedia data B- multimedia data E- multimedia data B- multimedia data D- multimedia data E... When determining the multimedia data subsequence associated with any object Um, the historical multimedia data sequence is split at a time interval of 3 hours, as shown in FIG. Figure 2e The part indicated by the dotted and solid lines can split the multimedia data operated between 07:00-14:00 into two multimedia data subsequences, namely, the multimedia data subsequence corresponding to 07:00-10:00: multimedia data C-multimedia data D-multimedia data A; and the multimedia data subsequence corresponding to 10:00-13:00: multimedia data B-multimedia data E-multimedia data B.

[0167] In this way, by setting a specified time length according to actual processing needs, the historical multimedia data sequence corresponding to each object is split, and a multimedia data subsequence consisting of each multimedia data operated by each object within the specified time length can be obtained, so that the multimedia data subsequence can reflect the operation interest of the corresponding object within the specified time length. Considering that the interests of the objects may be the same in the same time period, it is possible to a certain extent to divide the related multimedia data into a multimedia data subsequence.

[0168] Step 2013: The processing device generates corresponding nodes based on each multimedia data included in each obtained multimedia data subsequence, and connects the corresponding nodes according to the order in which each multimedia data is operated to obtain corresponding directed edges, wherein one multimedia data corresponds to one node in the graph structure.

[0169] After obtaining at least one multimedia data subsequence associated with each object, the processing device generates a corresponding node for each multimedia data included in each obtained multimedia data subsequence.

[0170] Furthermore, the processing device connects the nodes corresponding to the multimedia data according to the order in which each multimedia data in the multimedia data subsequence is operated to obtain corresponding directed edges, wherein the directed edge connects two nodes, and the direction of the directed edge is the same as the operation order of the two multimedia data corresponding to the two nodes.

[0171] See Figure 2fAs shown, it is a schematic diagram of the nodes corresponding to the generation of multimedia data in an embodiment of the present application. It is assumed that the multimedia data subsequences obtained by the processing device are: multimedia data C-multimedia data D-multimedia data A; multimedia data B-multimedia data E-multimedia data B; multimedia data D-multimedia data A-multimedia data B; multimedia data F-multimedia data A-multimedia data C; multimedia data B-multimedia data E-multimedia data D-multimedia data E. Then, the number of nodes is determined based on the total number of multimedia data types included in each multimedia data subsequence. For example, if each multimedia data subsequence includes multimedia data AF, the corresponding number of nodes is 6. Directed edges are then constructed between the nodes corresponding to the order of the multimedia data in each multimedia data subsequence. For example, in the multimedia data subsequence: multimedia data C - multimedia data D - multimedia data A, since multimedia data C is operated on after multimedia data D is operated on, there is a directed edge between the nodes corresponding to multimedia data C and D, from the node corresponding to multimedia data C to the node corresponding to multimedia data D. Furthermore, since multimedia data D is operated on after multimedia data A is operated on, there is a directed edge between the node corresponding to multimedia data D and the node corresponding to multimedia data A, pointing to the node corresponding to multimedia data A. Similarly, the construction of nodes and directed edges is completed.

[0172] Step 2014: The processing device completes the construction of the graph structure based on each node and each directed edge obtained.

[0173] Specifically, after completing the construction of nodes and directed edges based on at least one multimedia data subsequence associated with each object, the processing device completes the construction of a graph structure based on each node and the directed edges of each node, wherein the graph structure is composed of nodes and directed edges.

[0174] In this way, the relationship between each multimedia data in at least one multimedia data subsequence associated with each object is converted into a graph structure represented by nodes and directed edges, which is equivalent to establishing a relationship graph of global multimedia data.

[0175] Step 2015: The processing device performs the following operations for each directed edge in the graph structure: according to the direction of a directed edge, determines the order in which the corresponding two multimedia data are operated, and uses the number of operations in the first historical behavior data that matches the order in which the two multimedia data are operated as the edge weight of the corresponding directed edge.

[0176] Specifically, in order to characterize the differences in preferences for popular multimedia data or similar multimedia data, and to take into account the bias of each object towards different multimedia data in its first historical behavior data for multimedia data, the processing device can set a corresponding edge weight for each directed edge in the graph structure, wherein the value of the edge weight corresponding to the directed edge is positively correlated with the number of matches of the operation sequence of the two multimedia data corresponding to the directed edge in the first historical behavior data of each object. This allows subsequent sampling in each node of the graph structure to be more inclined to reflect the structural and homogeneous nature of the relationship network represented in the graph structure, wherein homogeneity means that the embedding vectors of nodes with similar distances in the graph structure should be as similar as possible, and structurality means that the embedding vectors of structurally similar nodes in the graph structure should be as similar as possible.

[0177] For example, assuming that the first historical behavior data of each object totals 100,000, and any directed edge 1 obtained represents the operation of multimedia data B after operating multimedia data A, and assuming that among the 100,000 first historical behavior data, there are 10,000 records of first operating multimedia data A and then operating multimedia data B, then the edge weight corresponding to directed edge 1 is considered to be 10,000.

[0178] Step 2016: The processing device uses the random walk algorithm of node2vec to sample each node in the graph structure based on the edge weight corresponding to each directed edge, and generates corresponding multimedia data sample sequences.

[0179] Specifically, the processing device samples each node included in the graph structure by random walk to obtain each node sequence, and generates corresponding each multimedia data sample sequence according to the correspondence between the node and the multimedia data, and determines the multimedia data input samples and multimedia data output samples used when training the classification model based on the multimedia sample sequence.

[0180] In the embodiment of the present application, the processing device adopts a random walk algorithm from node to vector node2vec, and balances the homogeneity and structure in the graph structure by adjusting parameters.

[0181] When the processing device samples the node2vec algorithm and determines the next node to jump to at a node in the random walk, the jump probability when jumping to different nodes can be determined using the following formula:

[0182] Assuming that node A is currently being processed, the probability of node A jumping to node X can be expressed as:

[0183] π Ax =α pq (t,x)·ω Ax

[0184] where ω Ax α is the edge weight of the directed edge AX. The more times the object continues to operate multimedia data X after operating multimedia data A in the first historical behavior data of each object, the greater the weight of the directed edge AX. pq (t,x) is a parameter, t represents the previous node that was wandered, X represents the next node that may be wandered to, α pq (t,x) is specifically defined as:

[0185]

[0186] d tx is the distance from node t to node X, representing the number of directed edges traversed when connecting node t and node X using the shortest distance. Parameters p and q jointly control the tendency of random walks. The smaller p, the better the network structure, and the smaller q, the better the network homogeneity. The values ​​of p and q are determined through testing. During the process of testing and determining the values ​​of p and q, by continuously adjusting the values ​​of p and q, a sequence of multimedia data samples obtained by random walks under different values ​​of p and q is obtained. Based on the multimedia data sample sequences, a classification model is trained to determine the corresponding association information feature vectors of the multimedia data. Based on the association information feature vectors of different multimedia data, recommended multimedia data is determined. The number of clicks on the recommended multimedia data is used as a measure of the recommendation effect to determine the values ​​of p and q that have the best recommendation effect.

[0187] In this way, with the help of the random walk method of node2vec, we can more fully consider the number of times popular multimedia data is operated, and take into account the differences in preferences for popular multimedia data or other multimedia data, so that the established multimedia data sample sequence can better reflect the click relationship between popular multimedia data and other multimedia data, and ensure that the subsequent associated information feature vectors generated corresponding to different multimedia data can reflect the internal relationship between multimedia data, which helps to improve the multimedia data recommendation effect.

[0188] Step 202: The processing device performs multiple rounds of iterative training on the classification model to be trained based on each multimedia data sample sequence until a preset convergence condition is met.

[0189] The processing device performs multiple rounds of iterative training on the classification model to be trained based on each obtained multimedia data sample sequence until a preset convergence condition is met, wherein the preset convergence condition may be that the classification model to be trained is trained based on all the obtained multimedia data sample sequences, and the preset convergence condition may also be that the loss value of the classification model is continuously lower than the set threshold for a number of times reaching a set value, and the loss value is calculated using a cross entropy loss function based on the difference between the predicted classification result output by the classification model to be trained and the classification result corresponding to the multimedia data output sample.

[0190] The following takes a training process as an example to illustrate the training process of the classification model to be trained:

[0191] See attached Figure 2g As shown, it is a flow chart of a training process in an embodiment of the present application.

[0192] Step 2021: The processing device determines the multimedia data input sample and the multimedia data output sample based on the currently obtained multimedia data sample sequence, and inputs the ID feature and label feature group of the multimedia data input sample into the classification model to be trained, obtains the associated information feature vector constructed corresponding to the multimedia data input sample, and obtains the predicted classification result output by the classification model to be trained.

[0193] After the processing device obtains each multimedia data sample sequence, the processing device performs the following operations for each multimedia data sample sequence: according to the arrangement order of each multimedia data included in a multimedia data sample sequence, the other multimedia data except the last multimedia data in a multimedia data sample sequence are obtained in sequence, and then the ID feature and label feature group corresponding to any obtained multimedia data are used as a multimedia data input sample, and another multimedia data that is operated on after any multimedia data is obtained, and the classification result corresponding to the other multimedia data is used as the multimedia data output sample corresponding to the multimedia data input sample.

[0194] It should be noted that, assuming that multimedia data X is used as a multimedia data input sample, and multimedia data Y that is subsequently operated on after multimedia data X is a corresponding multimedia data output sample, then for the multimedia data output sample, the processing device needs to determine the classification result of the multimedia data Y, wherein the classification result represents the attribution of the multimedia data Y in each classification category.

[0195] Specifically, when determining the classification result of multimedia data Y, the processing device may pre-set the classification result to include n classification categories, corresponding to n output information of the output layer (represented by O1, O2, O3, ..., On, respectively), so that each multimedia data item has a corresponding classification category. The classification category at different positions is a number between 0 and 1, representing the probability that the multimedia data item belongs to that classification category. When determining the classification result of multimedia data Y, the classification category corresponding to multimedia data Y is set to 1, and the classification information at other positions is set to 0, thereby obtaining the classification result corresponding to multimedia data Y.

[0196] For example, assuming that the total number of multimedia categories set is 6, corresponding to multimedia data A, F, and a multimedia data sample sequence is obtained as follows: multimedia data F - multimedia data A - multimedia data C. Based on this multimedia sequence, multimedia data F can be used as the multimedia data input sample, and multimedia data A can be used as the corresponding multimedia data output sample; and multimedia data A can be used as the multimedia data input sample, and multimedia data C can be used as the corresponding multimedia data output sample. When multimedia data A is used as the multimedia data output sample, the classification result of multimedia data A is (1, 0, 0, 0, 0, 0).

[0197] Furthermore, after the processing device determines the multimedia data input sample and the multimedia data output sample, the ID feature and label feature group of the multimedia data input sample are input into the classification model to be trained, and the associated information feature vector corresponding to the multimedia data input sample constructed by the classification model is obtained, and the predicted classification result output by the classification model to be trained is obtained, wherein the predicted classification result includes the n elements mapped after the classification model to be trained performs dimensional conversion processing based on the associated information feature vector, and the element value at each position represents the predicted probability of the classification category corresponding to the element position.

[0198] See Figure 2h As shown, it is a schematic diagram of the generation process of the associated information feature vector in the embodiment of the present application. Figure 2h , the process of obtaining the associated information feature vector corresponding to the multimedia data input sample by the processing device through the classification model to be trained is described:

[0199] Step 20211: The processing device inputs the ID features and label feature groups in the multimedia data sample into the classification model to be trained, maps the ID features into a first intermediate vector of a specified dimension through the classification model to be trained, and maps the label feature group into a second intermediate vector of a specified dimension through the classification model to be trained.

[0200] Specifically, the processing device inputs the ID features and label feature groups in the multimedia data sample into the classification model to be trained, obtains a first intermediate vector of a specified dimension after the classification model to be trained maps the ID features, and obtains a second intermediate vector of a specified dimension after the classification model to be trained maps the label feature group, wherein the dimensions of the first intermediate vector and the second intermediate vector are the same, and the value of the specified dimension is set according to actual processing needs.

[0201] It should be noted that in the embodiment of the present application, when the tag feature group contains only one tag feature, the tag feature can be directly mapped to the second intermediate vector of the specified dimension. However, when the tag feature group includes multiple tag features, it is necessary to perform mapping processing on each tag feature separately to obtain the corresponding vectors.

[0202] Specifically, when the label feature group of the multimedia data input sample includes multiple label features, the processing device splits the label feature group based on a preset delimiter through the classification model to be trained to obtain each label feature, and maps each label feature to a second intermediate sub-vector of a specified dimension corresponding to each of the respective labels, and then determines the elements at each position in each second intermediate sub-vector, and uses the average value of each element at the same position in each second intermediate sub-vector as the first new element at the corresponding position, and generates a second intermediate vector of the specified dimension corresponding to the label group based on the first new element corresponding to each position.

[0203] That is, in an embodiment of the present application, the label feature group of multimedia data may include at least one label feature, and different label features are separated by a preset delimiter. When the processing device constructs a second intermediate vector corresponding to the label feature group using the classification model to be trained, the processing device maps each label feature in the label feature group to a second intermediate sub-vector of a specified dimension using the classification model to be trained, and then uses the mean of the elements at the same position in each second intermediate sub-vector as the new element at the corresponding position, and constructs the second intermediate vector based on the obtained new elements at each position.

[0204] For example, assuming that the processing device maps the ID feature and the label feature group into a 128-dimensional vector through the classification model to be trained, and the label features in the label feature group are separated by semicolons, then the ID feature is mapped into a 128-dimensional first intermediate vector through the classification model to be trained, and the label features in the label feature group are split according to the semicolon through the classification model to be trained, and then each label feature is mapped into a 128-dimensional second intermediate sub-vector, and the mean of the elements at the same position in each 128-dimensional second intermediate sub-vector is used as the new element at the corresponding position, and then the second intermediate vector is obtained based on the obtained new elements.

[0205] For another example, assume that there are three 128-dimensional second intermediate sub-vectors, namely: (a1, a2, a3...a128), (b1, b2, b3...b128), and (c1, c2, c3...c128), then the resulting second intermediate vector is of the form ((a1+b1+c1) / 3, (a2+b2+c2) / 3, (a3+b3+c3) / 3, ..., (a128+b128+c128) / 3).

[0206] In possible implementations of the present application, the label feature group of multimedia data can be adjusted, and the adjustment method includes but is not limited to adding label features obtained manually or through training when operating multimedia data, or deleting inaccurate label features, or, while deleting inaccurate label features, adding label features obtained manually or through training when operating multimedia data.

[0207] Specifically, taking the presence of a new label in the label feature group of multimedia data as an example, when the processing device determines that there is a new label feature in the label feature group, the label feature group is split based on a preset delimiter through the classification model to be trained to obtain each label feature and each new label feature, and then the classification model to be trained is used to map each new label feature to a second intermediate new sub-vector of a specified dimension corresponding to each label feature, and determine the second intermediate sub-vector of a specified dimension corresponding to each label feature, and then determine the elements at each position in each second intermediate new sub-vector and each second intermediate sub-vector, and use the average value of each element at the same position in each second intermediate new sub-vector and each second intermediate sub-vector as the second new element at the corresponding position, and generate the second intermediate vector of the specified dimension corresponding to the modified label feature group based on the second new element corresponding to each position.

[0208] That is, when the label feature group of multimedia data changes, the processing device will regenerate the second intermediate sub-vectors corresponding to each label feature based on the label feature group of multimedia data, and generate a new second intermediate feature vector based on each second intermediate sub-vector.

[0209] In this way, the label feature group of multimedia data is scalable, the number of label features of multimedia data can be expanded, and the label features of multimedia data can be flexibly added or deleted. Moreover, no matter how many label features are included in the label feature group of multimedia data, the final result is a vector form of a specified dimension. In other words, when introducing new label features, there is no need to expand the feature dimension. They can be directly incorporated into the original label feature group and fused to obtain a second intermediate vector. This ensures that changes in the number of label features will not affect the processing architecture of the model, thereby improving the scalability of the model.

[0210] Step 20212: The processing device fuses the first intermediate vector and the second intermediate vector based on the first weight corresponding to the ID feature and the second weight corresponding to the label feature group through the classification model to be trained to obtain the association information feature vector constructed corresponding to the multimedia data input sample.

[0211] After the processing device constructs a first intermediate vector corresponding to the ID feature of the multimedia data input sample and a second intermediate vector corresponding to the label feature group of the multimedia data input sample through the classification model to be trained, the first intermediate vector and the second intermediate vector are fused based on the first weight corresponding to the ID feature and the second weight corresponding to the label feature group through the classification model to be trained to obtain the association information feature vector constructed by the classification model to be trained corresponding to the multimedia data input sample.

[0212] For example, assuming that in the classification model to be trained, the first weight is α0, the second weight is α1, the vector of the specified dimension is 128 dimensions, the first intermediate vector is in the shape of (x1, x2, x3...x128), and the second intermediate vector is in the shape of (y1, y2, y3...y128), then the obtained association information feature vector is in the shape of (α0*x1+α1*y1, α0*x2+α1*y2, α0*x3+α1*y3, ..., α0*x128+α1*y128).

[0213] It should be noted that in an embodiment of the present application, the processing device can set a classification model to store the constructed associated information feature vector corresponding to the multimedia data to a specified location, so that the associated information feature vector corresponding to each multimedia data of the construct can be obtained from the specified location later.

[0214] In this way, after the vectors corresponding to each label feature in the label special detection feature group are fused into a second intermediate vector, multiple label features are only corresponding to one second weight, and based on the first weight and the second weight in the classification model to be trained, the associated information feature vector of the classification model framework to be trained can be obtained, and the fused features of each multimedia data are converted into a vector space, which provides a basis for the subsequent determination of similar multimedia data.

[0215] Step 2022: The processing device calculates the loss value based on the difference between the predicted classification result and the multimedia data output sample, and uses gradient back propagation to adjust the first weight set for the corresponding ID feature and the second weight set for the corresponding label feature group.

[0216] After the processing device obtains the predicted classification result of the classification model to be trained based on the multimedia data input data output, it determines the element value difference between the predicted classification result and the classification result of the multimedia data output sample, and based on the determined element value difference, uses the cross entropy loss function to calculate the loss value, and then based on the loss value, uses the gradient back propagation method to adjust the first weight set for the corresponding ID feature and the second weight set for the corresponding label feature group.

[0217] In this way, only two weights need to be learned in the classification model, namely the first weight and the second weight. The parameters that need to be learned are very few and are not affected by the number of label features. This speeds up the training of the classification model while improving the scalability of the classification model.

[0218] See Figure 3 As shown in the figure, it is a schematic diagram of the multimedia data recommendation process in the embodiment of the present application. Figure 3 , the recommendation process of multimedia data in the embodiment of the present application is described:

[0219] Step 301: The processing device obtains a target association information feature vector corresponding to the target multimedia data in response to an operation instruction initiated on an operable page for the target multimedia data. The target association information feature vector is constructed using a classification model based on the following information: the identity ID feature of the target multimedia data and the corresponding first weight, and the label feature group of the target multimedia data and the corresponding second weight.

[0220] When the processing device determines that an operation instruction for target multimedia data has been initiated on an operable page, it obtains a target-related information feature vector corresponding to the target multimedia data, wherein the method for determining the initiation of the operation instruction for the target multimedia data can be that the processing device determines that the target object is on the operable page and operates on the target multimedia data, and the operation method includes but is not limited to clicking on the target multimedia data, or searching for keywords corresponding to the target multimedia data, etc., wherein the target object refers to a new account, an old account, or an unregistered temporary account that performs operations on the operable page, and the multimedia data can be such as video data, audio data, text data such as e-books, or emerging audio book data, etc.

[0221] It should be noted that, in the embodiment of the present application, generating the associated information feature vector corresponding to each multimedia data, and establishing the corresponding similar multimedia data set for each multimedia data, are operations before presenting the multimedia data on the operable page. That is to say, in the embodiment of the present application, the multimedia data that can be presented on the operable page are all pre-configured with the associated information feature vector and the similar multimedia data set.

[0222] In the embodiment of the present application, the target association information feature vector corresponding to the target multimedia data obtained by the processing device is constructed using a classification model based on the following information: the ID feature of the target multimedia data and the corresponding first weight, and the tag feature group of the target multimedia data and the corresponding second weight. Considering that the classification model can generate association information feature vectors for multimedia data before and after training, the process of the processing device obtaining the target association information feature vector may include the following two situations:

[0223] Case 1: When the classification model to be trained is trained using the target multimedia data, a target correlation information feature vector constructed by the classification model to be trained is obtained.

[0224] Specifically, when the target multimedia data is multimedia data that was published some time ago, historical operation records of different objects on the multimedia data can be obtained. Therefore, during the training phase of the classification model, the first historical behavior data of each object obtained includes the multimedia data. Therefore, a multimedia data sample sequence can be generated based on the operations on the multimedia data, so that the multimedia data can be used as a multimedia data input sample to train the classification model and obtain the associated information feature vector corresponding to the multimedia data constructed by the classification model. Therefore, the target associated information feature vector corresponding to the target multimedia data is constructed during the training process of the classification model to be trained.

[0225] It should be noted that since the classification model to be trained is in a periodic repeated training process, different associated information feature vectors may be generated for the same multimedia data in different training cycles. In the embodiment of the present application, only the latest associated information feature vector corresponding to the multimedia data is saved as the associated information feature vector corresponding to the multimedia data.

[0226] Case 2: Obtain a target association information feature vector constructed based on the ID feature and label feature group of the target multimedia data using a trained classification model.

[0227] Specifically, when the target multimedia data is newly released multimedia data, or multimedia data with a short release time, there may be no historical operation records for the multimedia data, and thus it is impossible to generate a multimedia data sample sequence including the multimedia data. Therefore, a trained classification model can be used, and the multimedia data can be used as input to obtain the associated information feature vector constructed by the classification model.

[0228] In this way, regardless of whether the target multimedia data has been operated, the processing device can obtain the target associated information feature vector corresponding to the target multimedia data, providing a basis for subsequent multimedia data recommendation based on the associated information feature vector.

[0229] After the processing device obtains the associated information feature vector corresponding to the multimedia data, it can establish similar multimedia data sets corresponding to each multimedia data based on the similarity between the associated information feature vectors of different multimedia data. The following describes the process of establishing similar multimedia data sets for target multimedia data as an example:

[0230] The processing device obtains a target association information feature vector corresponding to the target multimedia data, and based on the similarity between the target association information feature vector and the other association information feature vectors corresponding to each other multimedia data, screens out N multimedia data with the highest similarity to the target association information feature vector as similar multimedia data that meet the first set condition, and generates a similar multimedia data set corresponding to the target multimedia data based on each similar multimedia data.

[0231] Specifically, after the processing device obtains the target association information feature vector corresponding to the target multimedia data, it determines the similarity between the target association information feature vector and the association information feature vectors corresponding to each other multimedia data in the vector space constructed by the classification model, and then screens out the first similar multimedia data whose similarity with the target association information feature vector meets the first set condition.

[0232] It should be noted that the first set condition set by the processing device may be the N other related information feature vectors having the highest similarity to the target related information feature vector. The processing device may determine the similarity between the target related information feature vector and each other related information feature vector by using a Hierarchical Navigable Small World (HNSW) algorithm to calculate the similarity between different related information feature vectors, or by calculating the Euclidean distance between different related information feature vectors, or by calculating the cosine similarity between different related information feature vectors.

[0233] Furthermore, the processing device takes the N multimedia data corresponding to the screened N other associated information feature vectors as similar multimedia data that meet the first set condition with the target multimedia data, and generates a similar multimedia data set corresponding to the target multimedia data based on the N similar multimedia data, wherein each similar multimedia data in the similar multimedia data set can be arranged in descending order according to the similarity between the corresponding associated information feature vector and the target associated information feature vector.

[0234] It should be noted that in the embodiment of the present application, since the classification model is trained periodically, the associated information feature vector corresponding to the multimedia data is periodically updated and adjusted. Therefore, the processing device can periodically reconstruct the similar multimedia data set corresponding to each multimedia data based on the updated associated information feature vector corresponding to each multimedia data, wherein the reconstruction period of the similar multimedia data can be the same as the update period of the classification model. For example, after the training of the classification model is completed, the similar multimedia data is reconstructed immediately.

[0235] In this way, since the solution of this application relies on the ID features and tag features of the target multimedia data, for a newly-listed, non-behavior multimedia data, a corresponding set of similar multimedia data can still be obtained, thereby improving the interactivity with the multimedia data.

[0236] Step 302: The processing device obtains a similar multimedia data set established corresponding to the target multimedia data, and selects first similar multimedia data that meets a second set condition from the similar multimedia data set.

[0237] Specifically, the processing device can obtain M similar multimedia data in the similar multimedia data set in order of similarity from high to low with the target associated information feature vector, as the first similar multimedia data that meets the second set condition, where M is a preset integer, the similar multimedia data set includes similar multimedia data, and the similarity between the associated information feature vector corresponding to the similar multimedia data and the target associated information feature vector meets the first set condition.

[0238] After the processing device obtains the similar multimedia data set corresponding to the target multimedia data, it can select M similar multimedia data with the highest similarity to the target association information feature vector from the similar multimedia data set according to actual processing needs as the first similar multimedia data that meets the second set condition.

[0239] Optionally, the processing device may arbitrarily select M similar multimedia data from the similar multimedia data set as the first similar multimedia data according to actual processing needs.

[0240] In this way, since corresponding similar multimedia data sets are established in advance for each multimedia data, when the target multimedia data based on the object operation is subsequently recalled, the recalled multimedia data can be directly determined from the similar multimedia data set, thereby improving the recommendation efficiency of multimedia data.

[0241] Step 303: The processing device presents the first similar multimedia data that meets the second set condition as multimedia data to be recommended on the operable page.

[0242] After determining each first similar multimedia data that meets the second set condition between the target multimedia data and the target multimedia data, the processing device presents each first similar multimedia data as multimedia data to be recommended on the operable page.

[0243] In a possible embodiment of the present application, when displaying similar multimedia data, the processing device can selectively display the multimedia data to be recommended in the current operable page, or can selectively jump to other related pages of the target multimedia data and display the multimedia data to be recommended in the other related pages.

[0244] It should be noted that, when presenting similar multimedia data on an operational page, the processing device may, in response to an operation instruction initiated by an object on the operational page to display other similar multimedia data, continue to obtain M multimedia data from the similar multimedia data set as new first similar multimedia data, and present the new first similar multimedia data on the corresponding operational page. Optionally, after determining the first similar multimedia data, the processing device may determine, from the M first similar multimedia data, to present a portion on the operational page, and, in response to an operation instruction to display other similar multimedia data, present other content from the M first similar multimedia data on the operational page.

[0245] For example, in one possible presentation method, M takes a value of 3, and the similar multimedia data set includes 50 similar multimedia data. The processing device first obtains the similarity between the target associated information feature vector from the similar multimedia data set, and ranks 3 similar multimedia data with a similarity ranking of 1-3 from high to low as the first similar multimedia data of the target multimedia data, and presents them on the operable page. Then, in response to the operation instruction to display other similar multimedia data, the processing device continues to obtain 3 similar multimedia data with a similarity ranking of 4-6 from the similar multimedia data set, and presents them on the corresponding operable page.

[0246] For another example, in one possible presentation method, the value of M is 20, then the processing device can first obtain 3 first similar multimedia data from the 20 first similar multimedia data and present them on the operable page, and then continue to obtain 3 first similar multimedia data from the 20 first similar multimedia data in response to the operation instruction to display other similar multimedia data.

[0247] In an embodiment of the present application, the multimedia data to be recommended determined by the processing device may be of the same type as the target multimedia data. Optionally, when different types of multimedia data are mapped to the same vector space, the type of the multimedia data to be recommended may be different from the type of the target multimedia data.

[0248] For example, assuming that the target multimedia data is an e-book, when the classification model constructs a corresponding vector space based on multimedia data of the e-book type, the type of the multimedia data to be recommended is also determined to be an e-book.

[0249] For another example, assuming that the target multimedia data is an e-book, when the vector space constructed by the classification model includes various types of multimedia data, specifically, there are associated information feature vectors corresponding to e-book multimedia data in the vector space, there are associated information feature vectors corresponding to video multimedia data, there are associated information feature vectors corresponding to audio multimedia data, etc., then the similar multimedia data determined based on the e-book type multimedia data may be video type multimedia data, e-book type multimedia data, or audio type multimedia data.

[0250] In some possible embodiments of the present application, when multimedia data is presented on an operable page, the processing device can respond to a refresh operation on the operable page, filter out the object's multimedia data of interest based on the second historical behavior data of the object that logs into the operable page within a second specified time period, and obtain a set of similar multimedia data corresponding to the multimedia data of interest, and then filter out a specified number of second similar multimedia data from the set of similar multimedia data corresponding to the multimedia data of interest and present them on the operable page, and in response to an operation instruction initiated on the operable page for the target multimedia data, execute the step of obtaining a target association information feature vector corresponding to the target multimedia data.

[0251] Specifically, the processing device determines that a certain object (hereinafter referred to as object a) initiates a refresh operation on the operable page as an example for explanation. The processing device determines that object a enters the operable page, or the processing device determines that object a initiates a refresh operation on the operable page, or when the processing device determines that object a re-enters the operable page, it can be considered that a refresh operation on the operable page has been initiated. Then, the processing device filters out the multimedia data that object a is most interested in according to the second historical behavior data of object a logging into the operable page within the second specified time period, as the multimedia data of interest, and then determines a specified number of second similar multimedia data that are most similar to the multimedia data of interest according to the method of determining similar multimedia data proposed in this application, and then presents the second similar multimedia data on the operable page. Then, the processing device can respond to the operation initiated on the operable page for the target multimedia data and execute the operation defined in steps 301-303. Among them, the processing device can use the multimedia data that has been operated the most times in the second historical behavior data as the multimedia data of interest to object a.

[0252] In this way, the processing device can present recommended multimedia data at each stage when the object operates the operable page, so that the recommended multimedia data can be presented on each visual operable interface, thereby improving the object's usage experience while increasing the possibility of multimedia data being recalled, maximizing the exposure of multimedia data, and increasing the possibility of multimedia data being operated.

[0253] In summary, looking at the multimedia data recommendation process proposed in the embodiments of the present application, since the classification model is based on the ID features and label feature groups of the multimedia data, it is possible to construct the associated information feature vector corresponding to the multimedia data, and does not rely on the portrait features of the object. Therefore, the recommendation method proposed in the present application can be applied to multimedia data recommendations for new and old objects. Moreover, for the case where the target multimedia data is multimedia data that has not been operated, its corresponding similar multimedia data set can also be established in advance, so that the newly-listed multimedia data can also use the method proposed in the present application to recall the corresponding similar multimedia data or be recalled, thereby coping with the problem of cold start of multimedia data.

[0254] The multimedia data recommendation process will be further described below with reference to the accompanying drawings and specific recommendation scenarios.

[0255] The following is combined with Figure 4a-4c As shown, first, the presentation process of the operable page in the process of recommending multimedia data in the recommendation scenario in the embodiment of the present application is described. The following description only takes the triggering operation in response to object b on the operable page as an example for description.

[0256] See Figure 4a As shown, it is a schematic diagram of an operational page presented in an embodiment of the present application. The processing device determines that object b will access Figure 4a After the operable page is displayed, the second historical behavior data of object b in the past second specified time period (such as 1 week) is obtained, and based on the second historical behavior data, it is determined that object b has operated the multimedia data "Genius Catcher FM" the most times, so "Genius Catcher FM" is used as the interest multimedia data, and based on the interest multimedia data, a second similar multimedia data whose similarity meets the set conditions is determined, that is, the multimedia data "The Third Master's Teaching of His Son", and the relevant information of the determined multimedia data "The Third Master's Teaching of His Son" is presented on the operable page, and the display result is as follows: Figure 4a shown.

[0257] For further information, see Figure 4b As shown, it is a schematic diagram of an operable page further presented in an embodiment of the present application. Figure 4a Based on the Figure 4aIf the multimedia data "The Third Master's Teaching" recommended in the game generates a real-time operation behavior, "The Third Master's Teaching" is used as the target multimedia data, and the first similar multimedia data whose similarity with "The Third Master's Teaching" meets the set conditions are determined, such as "True or False Friends" and "The Road to Comedy", and the determined multimedia data "True or False Friends" and "The Road to Comedy" are presented on the operational page, and the display result is as follows Figure 4b shown.

[0258] See Figure 4c As shown, it is a schematic diagram of the playback interface presented for multimedia data in an embodiment of the present application, and the processing device determines that object b is in Figure 4a After initiating the operation of the multimedia data "Three Masters Teaching Their Sons" on the operation page, jump to Figure 4c The playback page of the multimedia data shown in the figure displays the text content of "The Third Master's Teachings" on the playback page, and the multimedia data recommended for "The Third Master's Teachings" are displayed on the current playback interface: "I Want to Travel", "I Am a Writer: The Four Great Classics", and "Interesting Stories about Life".

[0259] For further information, see Figure 4d , which is a flow chart of recommending electronic reading materials in the embodiment of this application, Figure 4d , the process of recommending e-books for a target e-book in a scenario where the target e-book has been previously operated is described:

[0260] When recommending e-books, the hardware environment of the processing device can be selectively configured as an Intel Xeon processor Gold 6133, 20 cores, 2.50GHz, and a graphics card of the processing device is equipped with a Tesla V100; and the software environment of the processing device is python2.7+tensorflow-gpu1.15.

[0261] Step 401: The processing device obtains ID features and tag feature groups of different electronic books.

[0262] It should be noted that, since electronic books are a type of multimedia data, the generation of ID features and tag feature groups of multimedia data has been described in detail in the above content and will not be repeated here.

[0263] Step 402: The processing device obtains the historical click behaviors of different accounts and obtains the click sequence of the electronic book.

[0264] For example, the processing device can obtain the operation logs within the past month, perform data cleaning on the operation logs, extract the account's click behavior on the e-books, and arrange the e-books in the order of the time when they were operated, to obtain an e-book sequence.

[0265] Step 403: The processing device generates input samples and output samples by random walk based on the click sequence of the electronic book, and obtains training samples for the classification model.

[0266] For example, after the processing device obtains the e-book click sequence of each account globally, it constructs a product relationship graph based on each e-book click sequence, and then generates input samples and output samples through random walk. The generation of training samples has been described in detail in the above process and will not be repeated here.

[0267] It should be noted that when generating input samples and output samples, the corresponding electronic readings need to be adjusted. For example, the electronic data used as the input sample should include the ID features and label feature groups of the electronic readings. Therefore, the ID features and label feature groups of the electronic readings need to be spliced ​​to obtain the input samples, and the classification results of the electronic readings used as sample outputs are constructed as the corresponding output samples.

[0268] Step 404: The processing device inputs the plurality of training samples into the preset model respectively to obtain the associated information feature vectors corresponding to the different electronic books.

[0269] Among them, the preset model is a classification model built according to actual processing needs. The classification model includes an input layer, an embedding layer, and an output layer. The embedding layer realizes feature mapping, mapping low-dimensional features to high-dimensional features of specified dimensions. The value of the specified dimension can be 128, 256, 512, etc.

[0270] Step 405: The processing device calculates the similarity between every two electronic books based on the associated information feature vectors of the electronic books, and obtains a similarity queue corresponding to each electronic book.

[0271] After the processing device obtains the associated information feature vector of the electronic book, it can selectively calculate the similarity of the associated information feature vector between each two electronic items, and obtain a corresponding similarity queue (similar multimedia data sequence) based on a certain number of the most similar electronic books, so that the content in the similarity queue can be directly obtained for recommendation in the future, thereby improving the recommendation efficiency.

[0272] Step 406: The processing device recalls other electronic books in a similar queue in response to the click behavior on the target electronic book initiated by the account.

[0273] After the processing device determines that the account initiates an operation on the target electronic book, it can directly obtain the electronic books in the similar column for recall.

[0274] Based on the technical solution proposed in this application, the applicant has achieved remarkable results during actual trial operation.

[0275] Referring to Tables 1 and 2 below, the solution proposed in this application corresponds to the "behavior-based recall" and "interest-based recall" methods in Table 1. The recall method proposed in this application can obtain the highest click-through rate in the experiment, which is better than the various recall methods used in the existing technology.

[0276] In addition, the technical solution proposed in this application also has a good application in the real-time click recommendation scenario. As shown in Table 2, compared with the existing real-time click recall method, the "data recommendation real-time click recall" method adopted by the solution of this application can obtain the maximum exposure, so that more multimedia data can be recalled and presented on the actionable page. Moreover, since the classification model in this application only relies on the ID and label feature group of the multimedia data, and does not rely on user features, for newly listed multimedia data sets, after determining the ID features and label feature groups, the technical solution proposed in the application can be used for recall, so it can deal with the problem of audio cold start. Among them, click unique visitors (UniqueVisitor, UV) represents the number of independent visitors who click on the exposed multimedia data under the corresponding recall method; the exposure number UV represents the number of independent visitors who present the corresponding multimedia data under the corresponding recall method; the UV click-through rate represents the percentage of click UV to exposure number UV.

[0277] Table 1

[0278]

[0279] Table 2

[0280]

[0281] See Figure 5 , which is a schematic diagram of the logical structure of a multimedia data recommendation device provided in an embodiment of the present application. The multimedia data recommendation device 500 may include:

[0282] A response unit 501 is configured to, in response to an operation instruction initiated on an operable page for target multimedia data, obtain a target-related information feature vector corresponding to the target multimedia data, where the target-related information feature vector is constructed using a classification model based on the following information: an ID feature of the target multimedia data and a corresponding first weight, and a tag feature group of the target multimedia data and a corresponding second weight, where the first weight and the second weight are iteratively adjusted during the training process of the classification model;

[0283] An acquisition unit 502 is configured to acquire a similar multimedia data set established corresponding to the target multimedia data, and to select first similar multimedia data satisfying a second set condition from the similar multimedia data set, wherein the similar multimedia data set includes similar multimedia data, and similarities between associated information feature vectors corresponding to the similar multimedia data and the target associated information feature vector satisfy the first set condition;

[0284] The presenting unit 503 is configured to present the first similar multimedia data meeting the second set condition as multimedia data to be recommended on the operable page.

[0285] Optionally, the apparatus further includes a training unit 504 for training the classification model in the following manner:

[0286] Based on first historical behavior data of each object within a first specified time period, a graph structure consisting of nodes and directed edges corresponding to the multimedia data is constructed, and edge weights of each directed edge are determined. Furthermore, a random walk method is used to sample nodes included in the graph structure based on the edge weights of each directed edge to generate corresponding sequences of multimedia data samples, wherein the edge weights corresponding to the directed edges are positively correlated with the number of times two multimedia data corresponding to the directed edges are operated on in order of their directions.

[0287] Based on each multimedia data sample sequence, the classification model to be trained is trained for multiple rounds of iterative training until the preset convergence conditions are met. In one training process, the following operations are performed:

[0288] Based on the currently obtained multimedia data sample sequence, determine the multimedia data input sample and the multimedia data output sample, and input the ID feature and label feature group of the multimedia data input sample into the classification model to be trained, obtain the associated information feature vector constructed corresponding to the multimedia data input sample, and obtain the predicted classification result output by the classification model to be trained;

[0289] Based on the difference between the predicted classification result and the multimedia data output sample, the loss value is calculated, and the first weight set for the corresponding ID feature and the second weight set for the corresponding label feature group are adjusted by using gradient back propagation.

[0290] Optionally, when the ID feature and label feature group in the multimedia data input sample are input into the classification model to be trained to obtain the associated information feature vector constructed corresponding to the multimedia data input sample, the training unit 504 is used to:

[0291] Inputting the ID features and the label feature group in the multimedia data sample into the classification model to be trained, mapping the ID features into a first intermediate vector of a specified dimension through the classification model to be trained, and mapping the label feature group into a second intermediate vector of a specified dimension through the classification model to be trained;

[0292] The first intermediate vector and the second intermediate vector are fused based on the first weight corresponding to the ID feature and the second weight corresponding to the label feature group through the classification model to be trained to obtain the association information feature vector constructed corresponding to the multimedia data input sample.

[0293] Optionally, when mapping the label feature group to a second intermediate vector of a specified dimension through the classification model to be trained, the training unit 504 is configured to:

[0294] Splitting the label feature group based on a preset separator using the classification model to be trained to obtain individual label features, and mapping each label feature to a second intermediate sub-vector of a corresponding specified dimension;

[0295] The elements at each position in each second intermediate sub-vector are determined respectively, and the average value of each element at the same position in each second intermediate sub-vector is used as the first new element at the corresponding position. Based on the first new element corresponding to each position, a second intermediate vector of the specified dimension corresponding to the label group is generated.

[0296] Optionally, the training unit 504 is further configured to:

[0297] When it is determined that there is a new label feature in the label feature group, the label feature group is split based on a preset separator using the classification model to be trained to obtain each label feature and each new label feature;

[0298] Mapping each newly added label feature to a second intermediate newly added sub-vector of a specified dimension respectively corresponding to each label feature through the classification model to be trained, and determining a second intermediate sub-vector of a specified dimension corresponding to each label feature;

[0299] Determine each second intermediate newly added sub-vector and the elements at each position in each second intermediate sub-vector respectively, take the average value of each second intermediate newly added sub-vector and each element at the same position in each second intermediate sub-vector as the second new element at the corresponding position, and generate a second intermediate vector of the specified dimension corresponding to the label feature group based on the second new elements corresponding to each position.

[0300] Optionally, when constructing a graph structure consisting of nodes and directed edges corresponding to multimedia data based on the first historical behavior data of each object within the first specified time period, the training unit 504 is configured to:

[0301] Obtaining first historical behavior data of each object within a first specified time period, and extracting a historical multimedia data sequence of each operation of each object within the first specified time period from the first historical behavior data, wherein each multimedia data in the historical multimedia data sequence is arranged in the order of being operated, and each multimedia data is associated with time information of being operated;

[0302] Splitting the historical multimedia data sequence corresponding to each object according to the specified time length to obtain at least one multimedia data subsequence associated with each object;

[0303] Generate corresponding nodes according to the multimedia data included in each obtained multimedia data subsequence, and connect the corresponding nodes according to the order in which the multimedia data are operated to obtain corresponding directed edges, wherein one multimedia data corresponds to one node in the graph structure;

[0304] Based on each node and each directed edge obtained, the graph structure is constructed.

[0305] Optionally, when a random walk method is used to sample nodes included in the graph structure based on the edge weights of each directed edge to generate corresponding multimedia data sample sequences, the training unit 504 is configured to:

[0306] For each directed edge in the graph structure, the following operations are performed: based on the direction of a directed edge, the order in which the corresponding two multimedia data are operated is determined, and the number of operations in the first historical behavior data that matches the order in which the two multimedia data are operated is used as the edge weight of the corresponding directed edge;

[0307] The node-to-vector random walk algorithm node2vec is adopted. Based on the edge weights corresponding to each directed edge, sampling is performed in each node of the graph structure to generate corresponding multimedia data sample sequences.

[0308] Optionally, when determining the multimedia data input sample and the multimedia data output sample based on the multimedia data sample sequence, the training unit 504 is configured to:

[0309] For each multimedia data sample sequence, perform the following operations:

[0310] Sequentially acquiring, according to the arrangement order of the multimedia data included in the multimedia data sample sequence, all multimedia data except the last multimedia data in the multimedia data sample sequence;

[0311] The ID feature and label feature group corresponding to any acquired multimedia data is used as a multimedia data input sample;

[0312] Another multimedia data that is subsequently operated after any multimedia data is obtained, and a classification result corresponding to the another multimedia data is used as a multimedia data output sample corresponding to the multimedia data input sample.

[0313] Optionally, when obtaining the target association information feature vector corresponding to the target multimedia data, the response unit 501 is configured to:

[0314] Obtaining a target association information feature vector constructed by the classification model to be trained when the classification model to be trained is trained using the target multimedia data; or

[0315] Obtain a target association information feature vector constructed based on the ID feature and label feature group of the target multimedia data using a trained classification model.

[0316] Optionally, before responding to an operation instruction initiated on the operable page for the target multimedia data, the responding unit 501 is further configured to:

[0317] Obtaining a target-related information feature vector corresponding to the target multimedia data, and based on similarities between the target-related information feature vector and other related information feature vectors corresponding to other multimedia data, screening out N multimedia data with the highest similarity to the target-related information feature vector as similar multimedia data that meet a first set condition, and generating a similar multimedia data set corresponding to the target multimedia data based on each similar multimedia data;

[0318] When first similar multimedia data meeting the second set condition is screened out from the similar multimedia data set, the acquiring unit 502 is configured to:

[0319] In descending order of similarity with the target association information feature vector, M similar multimedia data are obtained from the similar multimedia data set as first similar multimedia data that meet the second set condition, where M is a preset integer.

[0320] Optionally, the response unit 501 is further configured to:

[0321] In response to a refresh operation on the operable page, filtering out multimedia data of interest to the object based on second historical behavior data of the object logging into the operable page within a second specified time period, and obtaining a set of similar multimedia data corresponding to the multimedia data of interest;

[0322] From a set of similar multimedia data corresponding to the multimedia data of interest, a specified number of second similar multimedia data are screened out and presented on an operable page, and in response to an operation instruction initiated on the operable page for the target multimedia data, a step of obtaining a target association information feature vector corresponding to the target multimedia data is executed.

[0323] After introducing the multimedia data recommendation method and apparatus according to an exemplary embodiment of the present application, an electronic device according to another exemplary embodiment of the present application is introduced next.

[0324] Those skilled in the art will appreciate that various aspects of the present application can be implemented as systems, methods, or program products. Therefore, various aspects of the present application can be specifically implemented in the following forms: a complete hardware implementation, a complete software implementation (including firmware, microcode, etc.), or an implementation that combines hardware and software aspects, which may be collectively referred to herein as a "circuit," "module," or "system."

[0325] Based on the same inventive concept as the above method embodiment, an electronic device is also provided in the embodiment of the present application, Figure 6 , which is a schematic diagram of the hardware structure of an electronic device using an embodiment of the present application, wherein the electronic device 600 may include at least a processor 601 and a memory 602. The memory 602 stores program code, which, when executed by the processor 601, causes the processor 601 to perform any of the steps of the above-mentioned multimedia data recommendation method.

[0326] In some possible implementations, the computing device according to the present application may include at least one processor and at least one memory. The memory stores program code, and when the program code is executed by the processor, the processor performs the steps of multimedia data recommendation according to various exemplary implementations of the present application described above. For example, the processor may perform the following steps: Figure 3 Follow the steps shown in .

[0327] Refer to the following Figure 7 The computing device 700 according to this embodiment of the present application is described as follows. Figure 7 As shown, computing device 700 is implemented as a general-purpose computing device. Components of computing device 700 may include, but are not limited to, at least one processing unit 701, at least one storage unit 702, and a bus 703 connecting various system components (including storage unit 702 and processing unit 701).

[0328] Bus 703 represents one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, and a processor or local bus using any of a variety of bus architectures.

[0329] The storage unit 702 may include a readable medium in the form of a volatile memory, such as a random access memory (RAM) 7021 and / or a cache memory unit 7022 , and may further include a read-only memory (ROM) 7023 .

[0330] The storage unit 702 may also include a program / utility 7025 having a set (at least one) of program modules 7024, such program modules 7024 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

[0331] Computing device 700 can also communicate with one or more external devices 704 (e.g., a keyboard, pointing device, etc.), one or more devices that enable objects to interact with computing device 700, and / or any device that enables computing device 700 to communicate with one or more other computing devices (e.g., a router, modem, etc.). Such communication can occur via input / output (I / O) interface 705. Furthermore, computing device 700 can communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network such as the Internet) via network adapter 706. As shown, network adapter 706 communicates with other modules of computing device 700 via bus 703. It should be understood that, although not shown, other hardware and / or software modules can be used in conjunction with computing device 700, including but not limited to microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0332] Based on the same inventive concept as the above-mentioned method embodiment, various aspects of the key point detection method provided in this application can also be implemented in the form of a program product, which includes program code. When the program product is run on an electronic device, the program code is used to enable the electronic device to execute the steps of the multimedia data recommendation method according to various exemplary embodiments of the present application described above in this specification. For example, the electronic device can execute the following steps: Figure 3 Follow the steps shown in .

[0333] The program product may employ any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or component, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0334] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0335] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A multimedia data recommendation method, characterized in that: Applied to processing equipment, including: In response to an operation instruction initiated on an operable page for target multimedia data, a target-related information feature vector corresponding to the target multimedia data is obtained, wherein the target-related information feature vector is constructed using a classification model based on the following information: an identity ID feature of the target multimedia data and a corresponding first weight, and a tag feature group of the target multimedia data and a corresponding second weight, wherein the first weight and the second weight are iteratively adjusted during a training process of the classification model; the target-related information feature vector is used to classify a multimedia category of another multimedia data to be operated on subsequently to the target multimedia data; Obtaining a similar multimedia data set established corresponding to the target multimedia data, and screening out first similar multimedia data that meets a second set condition from the similar multimedia data set; the similar multimedia data set includes similar multimedia data, and similarity between a correlation information feature vector corresponding to the similar multimedia data and the target correlation information feature vector meets the first set condition; presenting the first similar multimedia data satisfying the second set condition as multimedia data to be recommended on the operable page; In which, the target association information feature vector is constructed by the classification model during training or the classification model after training; when obtaining the target association information feature vector, the ID feature and the label feature group of the target multimedia data are input into the corresponding classification model, and the ID feature is mapped to a first intermediate vector of a specified dimension through the classification model, and the label feature group is mapped to a second intermediate vector of a specified dimension through the classification model; and, based on the first weight corresponding to the ID feature and the second weight corresponding to the label feature group, the first intermediate vector and the second intermediate vector are fused through the classification model to obtain the association information feature vector constructed corresponding to the target multimedia data.

2. The method according to claim 1, wherein The classification model is trained in the following manner, including: Based on first historical behavior data of each object within a first specified time period, a graph structure consisting of nodes and directed edges corresponding to the multimedia data is constructed, and an edge weight of each directed edge is determined. Furthermore, a random walk method is used to sample nodes included in the graph structure based on the edge weights of each directed edge to generate corresponding sequences of multimedia data samples, wherein the edge weight corresponding to the directed edge is positively correlated with the number of times two multimedia data corresponding to the directed edge are operated on in order of their directions. Based on the multimedia data sample sequences, the classification model to be trained is subjected to multiple rounds of iterative training until a preset convergence condition is met; wherein, during one training process, the following operations are performed: Determining a multimedia data input sample and a multimedia data output sample based on a currently obtained multimedia data sample sequence, and inputting an ID feature and a label feature group of the multimedia data input sample into a classification model to be trained, obtaining an associated information feature vector constructed corresponding to the multimedia data input sample, and obtaining a predicted classification result output by the classification model to be trained; Based on the difference between the predicted classification result and the multimedia data output sample, a loss value is calculated, and a first weight corresponding to the ID feature setting and a second weight corresponding to the label feature group setting are adjusted by using gradient back propagation.

3. The method according to claim 2, wherein The step of inputting the ID feature and the label feature group in the multimedia data input sample into the classification model to be trained to obtain the associated information feature vector corresponding to the multimedia data input sample comprises: Inputting the ID features and the label feature group in the multimedia data input sample into a classification model to be trained, mapping the ID features into a first intermediate vector of a specified dimension through the classification model to be trained, and mapping the label feature group into a second intermediate vector of a specified dimension through the classification model to be trained; The first intermediate vector and the second intermediate vector are fused based on the first weight corresponding to the ID feature and the second weight corresponding to the label feature group by the classification model to be trained to obtain an associated information feature vector constructed corresponding to the multimedia data input sample.

4. The method according to claim 3, wherein Mapping the label feature group into a second intermediate vector of a specified dimension by the classification model to be trained includes: Splitting the label feature group based on a preset separator using the classification model to be trained to obtain individual label features, and mapping each label feature into a second intermediate sub-vector of a corresponding specified dimension; The elements at each position in each second intermediate sub-vector are determined respectively, and the average value of each element at the same position in each second intermediate sub-vector is used as the first new element at the corresponding position, and based on the first new elements corresponding to each position, a second intermediate vector of the specified dimension corresponding to the label feature group is generated.

5. The method according to claim 4, wherein Also includes: When it is determined that there is a newly added label feature in the label feature group, the label feature group is split based on a preset separator by the classification model to be trained to obtain each label feature and each newly added label feature; Mapping each newly added label feature to a second intermediate newly added sub-vector of a specified dimension respectively corresponding to each label feature through the classification model to be trained, and determining a second intermediate sub-vector of the specified dimension respectively corresponding to each label feature; Determine each second intermediate newly added sub-vector and each element at each position in each second intermediate sub-vector, and use the average value of each element at the same position in each second intermediate newly added sub-vector and each second intermediate sub-vector as the second new element at the corresponding position, and generate a second intermediate vector of the specified dimension corresponding to the label feature group based on the second new elements corresponding to each position.

6. The method according to any one of claims 2 to 5, characterized in that The step of constructing a graph structure consisting of nodes and directed edges corresponding to multimedia data based on the first historical behavior data of each object within the first specified time period includes: Obtaining first historical behavior data of each object within a first specified time period, and extracting a historical multimedia data sequence of each object's respective operations within the first specified time period from the first historical behavior data, wherein each multimedia data in the historical multimedia data sequence is arranged in the order in which it was operated, and each multimedia data is associated with time information of the operation; Splitting the historical multimedia data sequences corresponding to the respective objects according to the specified time lengths to obtain at least one multimedia data subsequence associated with the respective objects; Generate corresponding nodes according to the multimedia data included in each obtained multimedia data subsequence, and connect the corresponding nodes according to the order in which the multimedia data are operated to obtain corresponding directed edges, wherein one multimedia data corresponds to one node in the graph structure; Based on the nodes and the directed edges obtained, the construction of the graph structure is completed.

7. The method according to any one of claims 2 to 5, wherein: The random walk method is used to sample each node included in the graph structure based on the edge weight of each directed edge to generate corresponding multimedia data sample sequences, including: For each directed edge in the graph structure, performing the following operations: determining, based on the direction of a directed edge, a sequence in which two corresponding multimedia data items are operated, and using a number of operations in the first historical behavior data that matches the sequence in which the two multimedia data items are operated as an edge weight of the corresponding directed edge; A random walk algorithm from node to vector node2vec is adopted, and based on the edge weights corresponding to the directed edges, sampling is performed in each node of the graph structure to generate corresponding multimedia data sample sequences.

8. The method according to any one of claims 2 to 5, wherein: The determining of the multimedia data input sample and the multimedia data output sample based on the multimedia data sample sequence includes: For each multimedia data sample sequence, perform the following operations: Sequentially acquiring, according to the arrangement order of the multimedia data included in a multimedia data sample sequence, the multimedia data except the last multimedia data in the multimedia data sample sequence; The ID feature and label feature group corresponding to any acquired multimedia data is used as a multimedia data input sample; Another multimedia data that is subsequently operated after the any multimedia data is acquired, and a classification result corresponding to the another multimedia data is used as a multimedia data output sample corresponding to the multimedia data input sample.

9. The method according to claim 1, wherein The obtaining of the target association information feature vector corresponding to the target multimedia data includes: obtaining a target association information feature vector constructed by the classification model to be trained when the classification model to be trained is trained using the target multimedia data; or A target association information feature vector constructed based on the ID feature and label feature group of the target multimedia data using a trained classification model is obtained.

10. The method according to any one of claims 1 to 5 and 9, characterized in that: Before responding to the operation instruction initiated on the operable page for the target multimedia data, the method further includes: Obtaining a target-related information feature vector corresponding to the target multimedia data, and based on similarities between the target-related information feature vector and other related information feature vectors corresponding to other multimedia data, screening out N multimedia data with the highest similarity to the target-related information feature vector as similar multimedia data that meet a first set condition, and generating a similar multimedia data set corresponding to the target multimedia data based on each similar multimedia data; The step of screening out first similar multimedia data that meets a second set condition from the similar multimedia data set includes: In descending order of similarity with the target association information feature vector, M similar multimedia data are obtained from the similar multimedia data set as first similar multimedia data that meet the second set condition, where M is a preset integer.

11. The method according to any one of claims 1 to 5 and 9, characterized in that: Also includes: In response to a refresh operation on the operable page, filtering out multimedia data of interest to the object based on second historical behavior data of the object that logged into the operable page within a second specified time period, and obtaining a set of similar multimedia data corresponding to the multimedia data of interest; From a set of similar multimedia data corresponding to the multimedia data of interest, a specified number of second similar multimedia data are screened out and presented on the operable page, and in response to an operation instruction initiated on the operable page for the target multimedia data, a step of obtaining a target association information feature vector corresponding to the target multimedia data is performed.

12. A multimedia data recommendation device, characterized in that: include: a response unit, configured to, in response to an operation instruction initiated on an operable page for target multimedia data, obtain a target-related information feature vector corresponding to the target multimedia data, wherein the target-related information feature vector is constructed using a classification model based on the following information: an identity ID feature of the target multimedia data and a corresponding first weight, and a tag feature group of the target multimedia data and a corresponding second weight, wherein the first weight and the second weight are two parameters and are iteratively adjusted during the training process of the classification model; the target-related information feature vector is used to classify a multimedia category of another multimedia data to be operated on subsequently to the target multimedia data; an acquisition unit, configured to acquire a similar multimedia data set established corresponding to the target multimedia data, and screen out first similar multimedia data that meets a second set condition from the similar multimedia data set, wherein the similar multimedia data set includes similar multimedia data, and similarity between a correlation information feature vector corresponding to the similar multimedia data and a target correlation information feature vector satisfies the first set condition; a presenting unit, configured to present the first similar multimedia data satisfying the second set condition as multimedia data to be recommended on the operable page; In which, the target association information feature vector is constructed by the classification model during training or the classification model after training; when obtaining the target association information feature vector, the ID feature and the label feature group of the target multimedia data are input into the corresponding classification model, and the ID feature is mapped to a first intermediate vector of a specified dimension through the classification model, and the label feature group is mapped to a second intermediate vector of a specified dimension through the classification model; and, based on the first weight corresponding to the ID feature and the second weight corresponding to the label feature group, the first intermediate vector and the second intermediate vector are fused through the classification model to obtain the association information feature vector constructed corresponding to the target multimedia data.

13. The device according to claim 12, wherein The device further includes a training unit for training the classification model in the following manner: Based on first historical behavior data of each object within a first specified time period, a graph structure consisting of nodes and directed edges corresponding to the multimedia data is constructed, and an edge weight of each directed edge is determined. Furthermore, a random walk method is used to sample nodes included in the graph structure based on the edge weights of each directed edge to generate corresponding sequences of multimedia data samples, wherein the edge weight corresponding to the directed edge is positively correlated with the number of times two multimedia data corresponding to the directed edge are operated on in order of their directions. Based on the multimedia data sample sequences, the classification model to be trained is subjected to multiple rounds of iterative training until a preset convergence condition is met; wherein, during one training process, the following operations are performed: Determining a multimedia data input sample and a multimedia data output sample based on a currently obtained multimedia data sample sequence, and inputting an ID feature and a label feature group of the multimedia data input sample into a classification model to be trained, obtaining an associated information feature vector constructed corresponding to the multimedia data input sample, and obtaining a predicted classification result output by the classification model to be trained; Based on the difference between the predicted classification result and the multimedia data output sample, a loss value is calculated, and a first weight corresponding to the ID feature setting and a second weight corresponding to the label feature group setting are adjusted by using gradient back propagation.

14. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores program codes, and when the program codes are executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 11.

15. A computer-readable storage medium, characterized in that It includes program codes. When the program product is run on an electronic device, the program codes are used to enable the electronic device to execute the steps of the method according to any one of claims 1 to 11.

Images