A training method, a recommendation method and a device

Abstract

The present disclosure provides a training method, a recommendation method and a device, the method comprising: obtaining graph structure sample data based on user basic information and heterogeneous behavior information of a user for a multimedia product, the graph structure sample data comprising node data of a plurality of node samples, a plurality of behavior sample data and relationship information of the node data, and training a behavior prediction model based on the graph structure sample data. The method can effectively extract multiple behavior preference prediction information of a user for a target type multimedia product, achieve deep modeling of high-order heterogeneous behavior, and accurately recommend multimedia products.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent recommendation, and in particular to a training method and a recommendation method and device. BACKGROUND

[0002] In a video recommendation scenario, there can be a large number of sports video product users who are usually not interested in video and variety video products. A video playing platform hopes to recommend interested video and variety video products to similar users through video recommendation, so that similar users can stay on the video playing platform and improve the user retention of the video playing platform.

[0003] In related technologies, a collaborative filtering learning mode can be used to recommend video and variety video products to users. However, the model of this recommendation mode has weak expression ability and cannot accurately recommend video and variety video products to users, causing the problem of difficulty in recommending video and variety video products. SUMMARY

[0004] According to an aspect of the present application, a training method is provided, comprising:

[0005] Based on user basic information and user heterogeneous behavior information for multimedia products, a graph structure sample data is obtained, the graph structure sample data comprising node data of a plurality of node samples, a plurality of behavior sample data and relationship information of the node data, the plurality of node samples comprising a user sample and a multimedia product sample, and the relationship information of the node data comprising a heterogeneous behavior relationship between the user sample and the multimedia product sample;

[0006] A behavior prediction model is trained based on the graph structure sample data, the behavior prediction model comprising a graph neural network and a behavior preference prediction layer, the graph neural network being configured to obtain node update data of a node sample based on node data and behavior sample data of adjacent node samples associated with the same node sample, and the behavior preference prediction layer being configured to determine multiple behavior preference prediction information of multiple user samples for different multimedia product samples based on multiple behavior sample data, node update data of multiple user samples and node update data of multiple multimedia product samples.

[0007] According to another aspect of the present application, a recommendation method is provided, comprising:

[0008] Target user data, target behavior data and a plurality of multimedia product data are input into the behavior prediction model trained by the exemplary embodiments of the present application;

[0009] obtain, by using the behavior prediction model, multiple behavior preference prediction information of the target user for different multimedia product samples from multiple user samples for different multimedia product samples;

[0010] recommend the target multimedia product if the target behavior preference information of the target user for the target multimedia product meets a recommendation condition.

[0011] According to another aspect of the present disclosure, a training device is provided, comprising:

[0012] an obtaining module configured to obtain graph structure sample data based on user basic information and heterogeneous behavior information of users for multimedia products, the graph structure sample data comprising node data of multiple node samples, multiple behavior sample data, and relationship information of the node data, the multiple node samples comprising user samples and multimedia product samples, the multiple behavior sample data comprising multiple, and the relationship information of the node data comprising heterogeneous behavior relationships between the user samples and the multimedia product samples;

[0013] a training module configured to train a behavior prediction model based on the graph structure sample data, the behavior prediction model comprising a graph neural network and a behavior preference prediction layer, the graph neural network being configured to obtain node update data of a node sample based on node data and behavior sample data of adjacent node samples associated with the same node sample, and the behavior preference prediction layer being configured to determine multiple behavior preference prediction information of multiple user samples for different multimedia product samples based on multiple behavior sample data, node update data of multiple user samples, and node update data of multiple multimedia product samples.

[0014] According to another aspect of the present disclosure, a recommendation device is provided, comprising:

[0015] an inference module configured to input target user data, target behavior data, and multiple multimedia product data into the behavior prediction model trained by the exemplary embodiments of the present disclosure, and obtain, by using the behavior prediction model, multiple behavior preference prediction information of the target user for different multimedia product samples from multiple user samples for different multimedia product samples;

[0016] a recommendation module configured to recommend a target multimedia product if target behavior preference information of the target user for the target multimedia product meets a recommendation condition.

[0017] According to another aspect of the present disclosure, an electronic device is provided, comprising:

[0018] a processor; and

[0019] a memory storing programs;

[0020] The program includes instructions that, when executed by the processor, cause the processor to perform the method according to the exemplary embodiments of the present disclosure.

[0021] According to another aspect of the present disclosure, a non-transitory computer-readable storage medium is provided, and the non-transitory computer-readable storage medium stores computer instructions for causing the computer to perform the method according to the exemplary embodiments of the present disclosure.

[0022] In one or more technical solutions provided in the exemplary embodiments of the present disclosure, the graph structure sample data can be obtained based on the basic user information and the heterogeneous behavior information of the user for the multimedia product, so that the graph structure sample data includes node data of a plurality of node samples, a plurality of behavior sample data and relationship information of the node data. It can be seen that the exemplary embodiments of the present disclosure can introduce high-order heterogeneous behavior of the user for the multimedia product into the graph structure sample data, and form a heterogeneous graph with the user sample and the multimedia product sample as nodes, and the heterogeneous behavior between the user sample and the multimedia product sample as edges.

[0023] When the behavior prediction model is trained based on the multi-relation graph structure data, the graph neural network is used to determine the node update data of the corresponding target node sample based on the node data of the neighbor node sample adjacent to each target node sample, and the behavior preference prediction layer is used to determine the multiple behavior preference prediction information of the user sample for the multimedia product sample based on the multiple behavior sample data and the node update data of the multiple node samples. It can be seen that the behavior prediction model not only can fuse the heterogeneous behavior information into the user and the multimedia product, but also can consider the high-order connection between the user, the multimedia product and the user behavior through the information transmission process, so as to fully analyze the sparse behavior information of the user for the target type multimedia product, thereby effectively extracting the multiple behavior preference prediction information of the user for the target type multimedia product, so as to achieve the purpose of deep modeling of the high-order heterogeneous behavior.

[0024] Moreover, the graph neural network is used to determine the node update data of the corresponding node sample based on the node data of the neighbor node sample adjacent to each node sample, without fusing the node data of the node sample itself, so as to reduce the model parameter quantity, make the behavior prediction model lightweight, and improve the behavior preference prediction speed of the behavior prediction model. Meanwhile, the self-connection of the node sample is captured by means of the interlayer connection operation of the graph neural network, so that the node data loss problem caused by fusing only the node data of the neighbor node sample can be avoided. Therefore, the behavior preference prediction layer in the example embodiment of the disclosure can accurately determine the multiple behavior preference prediction information of the user sample for the multimedia product sample based on the multiple behavior sample information and the node update information of the multiple node samples, and can make sufficient preparation for multimedia product recommendation.

[0025] In the multimedia product recommendation scene, the method in the example embodiment of the disclosure can input the target user data, the target behavior data and the multiple multimedia product data into the trained behavior prediction model, thereby obtaining the multiple behavior preference information of the target user for the multiple multimedia products, and then determining whether the target behavior preference information of the target user for the target multimedia product meets the recommendation condition based on the collaborative filtering mode. If the recommendation condition is met, it means that the target user has a relatively high possibility to apply the target behavior to the target multimedia product, and the target multimedia product can be recommended.

[0026] It can be seen that, in the method in the example embodiment of the disclosure, if the target user has relatively sparse behavior information for a certain type of multimedia product due to personal daily preference problems, the target user data, the target behavior data and the multiple multimedia product data can be input into the trained behavior prediction model, so as to analyze the sparse behavior information of the target user for the multimedia product by using the behavior prediction model from the multiple behavior preference prediction information of the multiple user samples for different multimedia product samples, that is, the high-order connection among the user samples, the multimedia product samples and the behavior samples, thereby obtaining the multiple behavior preference information of the target user for the multiple multimedia products, and then accurately recommending the target multimedia product with relatively sparse behavior information to the target user on the basis of this, so as to increase the attraction of the recommended target multimedia data to the target user, and improve the user retention rate of the multimedia playing platform. BRIEF DESCRIPTION OF DRAWINGS

[0027] In the following description of the example embodiments with reference to the accompanying drawings, more details, features and advantages of the disclosure are disclosed, and in the drawings:

[0028] Figure 1 A flowchart of a training method of the example embodiment of the disclosure is shown;

[0029] Figure 2A flowchart of a recommendation method according to an example embodiment of the present disclosure is shown.

[0030] Figure 3 A schematic block diagram of a training device according to an example embodiment of the present disclosure is shown.

[0031] Figure 4 A schematic block diagram of a recommendation device according to an example embodiment of the present disclosure is shown.

[0032] Figure 5 A schematic block diagram of a chip according to an example embodiment of the present disclosure is shown.

[0033] Figure 6 A structural block diagram of an example electronic device that can be used to implement embodiments of the present disclosure is shown. DETAILED DESCRIPTION

[0034] Embodiments of the present disclosure will be described in more detail with reference to the drawings. While certain embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be embodied in various forms and should not be interpreted as being limited to the embodiments set forth herein; rather, these embodiments are provided so that the present disclosure can be more thoroughly and completely understood. It should be understood that the drawings and embodiments of the present disclosure are only for illustrative purposes and are not intended to limit the scope of protection of the present disclosure.

[0035] It should be understood that the various steps in the method embodiments of the present disclosure can be performed in different orders and / or in parallel. In addition, the method embodiments can include additional steps and / or omit performing the steps shown. The scope of the present disclosure is not limited in this respect.

[0036] The term "comprising" and variations thereof as used herein are used inclusively, i.e., "comprising, but not limited to." The term "based on" is "based at least in part on." The term "one embodiment" means "at least one embodiment." The term "another embodiment" means "at least one additional embodiment." The term "some embodiments" means "at least some embodiments." Related definitions of other terms will be given in the description below. It should be noted that the concepts "first," "second," etc. mentioned in the present disclosure are only used to distinguish different devices, modules or units, and are not intended to limit the order or interdependence of the functions performed by these devices, modules or units.

[0037] It should be noted that the modification "one" or "more" mentioned in the present disclosure is illustrative and not limiting, and those skilled in the art should understand that, unless otherwise explicitly stated in the context, it should be understood as "one or more."

[0038] Before introducing the embodiments of the present disclosure, the related terms involved in the embodiments of the present disclosure are first explained as follows:

[0039] Collaborative filtering (CF) is a recommendation algorithm designed based on user behavior, which finds some similarity (similarity between users or similarity between items) through group behavior, and makes decisions and recommendations for users through similarity. In a literal sense, collaborative filtering includes two operations of collaboration and filtering. Collaboration is to collect all user feedback, evaluation, etc. (interact with the website continuously. And filtering is the process of obtaining information through collaboration, which can filter out the items of interest to users from a large number of items.

[0040] Graph Neural Networks (GNNs) is a graph structure-based deep learning method that applies the predictive power of deep learning to rich data structures that describe objects and their relationships as points connected by lines in a graph. In GNN, data points are called nodes, which are connected to elements represented mathematically by lines called edges, so that machine learning algorithms can make useful predictions at the node, edge, or entire graph level.

[0041] Graph Convolutional Networks (GCN) is a convolutional neural network that can directly act on a graph and utilize its structural information. The main idea of GCN is that for each node, all its neighbors and the feature information it contains need to be considered.

[0042] Adjacency Matrix is a matrix that represents the adjacency relationship between vertices in a graph, which is an n-order square matrix (n is the number of vertices). Its logical structure is divided into two parts: vertex set and edge set. All vertex data in the graph can be stored in a one-dimensional array; the relationship (edge or arc) between vertices can be stored in a two-dimensional array, which is called an adjacency matrix.

[0043] Degree is a diagonal matrix, and the elements on the diagonal are the degrees of each vertex in the graph. The degree of a vertex represents the number of edges associated with the vertex. In an undirected graph, the degree of vertex vi is d(vi) = N(i), where N(i) represents the number of edges. In a directed graph, the degree of vertex vi is divided into the out-degree and in-degree of vertex vi, which are the number of directed edges going out of vertex vi and the number of directed edges entering vertex vi.

[0044] Heterogeneous behavior refers to the multiple interaction relationships between users and goods, which has the behavior heterogeneity characteristic, and each behavior contains different semantics. For example, different behaviors generally reflect different user preference degrees, clicks usually only mean that users have a preliminary and relatively vague interest in goods, and purchases represent a strong user preference.

[0045] Heterogeneous graph, also known as heterogeneous graph, refers to a graph in which the types of nodes and edges are not only one type. The sum of the number of categories is greater than 2, which can be called a heterogeneous graph.

[0046] In the video recommendation scene, there may be a large number of sports video product users who are usually not interested in video, variety video products. The video playing platform hopes to recommend interested video, variety video products to similar users through video recommendation, so that similar users can stay on the video playing platform for a long time, and improve the user retention of the video playing platform.

[0047] At present, the video recommendation system usually adopts the traditional collaborative filtering learning mode to focus on one type of user behavior, and simultaneously models the user behavior of this type based on a neural network model, so as to predict the preference information of the target user for the type of user behavior by using the neural network model. In the scene where the user behavior of the user for the video, variety video product is sparse, it is difficult to effectively learn the preference information of the user for the video, variety video, resulting in difficulty in recommending the video, variety video to the user.

[0048] The inventors found that in the video playing process, the same user has multiple behavior characteristics for the target video product, such as browsing and staying, jumping from the source, clicking, playing time, collecting, commenting, completing playing, and downloading. These heterogeneous behaviors can be defined as heterogeneous behaviors, which provide valuable information for user preference learning. However, the related video recommendation system only uses one type of user behavior of the user and the video product for video recommendation, and does not perform in-depth analysis on the heterogeneous behavior of the user, thereby greatly reducing the expression ability of the neural network model, and making it difficult for the video recommendation system to accurately recommend the video, variety video product with sparse user behavior to the user.

[0049] To solve the above problems, the exemplary embodiments of the present disclosure provide a training method and a recommendation method, which can fully analyze the heterogeneous behavior information of the user, obtain the multiple behavior preference information of the user for the target multimedia product, and accurately recommend the multimedia product to the user by using the multiple behavior preference information of the user for the target multimedia product.

[0050] In a multimedia playing platform, multimedia products can be divided into multiple types according to the content fields involved. The user usually has more user behaviors for one or more types of multimedia products, so that the user behaviors of these types of multimedia products are more dense, and has less user behaviors for another type or multiple types of multimedia products, so that the user behaviors of these types of multimedia products are more sparse.

[0051] The exemplary embodiments of the present disclosure can use a training method to mine various user heterogeneous behavior information in a multimedia playing platform, model the heterogeneous behavior of video products, analyze the high-order relationship among users, multimedia products and behaviors in the user heterogeneous behavior information, and determine the multiple behavior preference prediction information of a user sample for a multimedia product sample. On this basis, the behavior prediction model is used to obtain the multiple preference information of a target user for a behavior sparse multimedia product, so as to accurately recommend such multimedia products and improve the user retention rate of the multimedia playing platform.

[0052] The multimedia product involved in the method of the exemplary embodiments of the present disclosure can be a video product, an audio product, a graphic text product, etc. Taking a video playing platform as an example, the video content involved in the video playing platform can divide the types of video products into film and television, variety, sports, science and education, etc. Among the users of the video playing platform, some users prefer sports video products and often generate user behaviors for sports video products, and are not interested in film and television, variety and science and education video products. It can be seen that for such users, the behavior information of sports video products is dense information, and the behavior information of film and television and variety video products is sparse information.

[0053] The method of the exemplary embodiments of the present disclosure can analyze the high-order relationship among users, video products and behaviors in the user heterogeneous behavior information by training a behavior prediction model, so as to determine the multiple behavior preference prediction information of a user sample for a video product. It can be seen that the method of the exemplary embodiments of the present disclosure can accurately recommend film and television and variety video products for a target user and improve the user retention rate of the video playing platform.

[0054] The training method and the recommendation method of the exemplary embodiments of the present disclosure can be applied to an electronic device or a chip of an electronic device. The electronic device can be various computer devices, such as a server, a mobile phone, a tablet computer, a wearable device, a vehicle-mounted device, a notebook computer, an ultra-mobile personal computer (UMPC), a netbook, a Personal Digital Assistant (PDA), and a wearable device based on augmented reality (AR) and / or virtual reality (VR) technology, etc.

[0055] Exemplarily, when the electronic device is a wearable device, the wearable device can also be a general term of devices developed by applying wearable technology to intelligent design of daily wear, such as glasses, gloves, watches, clothing, and shoes. The wearable device is a portable device directly worn on the body or integrated into the clothes or accessories of the user. The wearable device is not only a hardware device, but also a device that realizes powerful functions through software support and data interaction and cloud interaction. The general wearable smart device includes a device with full functions and large size, which can realize complete or partial functions without relying on a smart phone, such as a smart watch or smart glasses, and a device that focuses on only one type of application function and needs to cooperate with other devices such as a smart phone, such as various smart bracelets and smart jewelry for monitoring vital signs.

[0056] Figure 1 A flowchart of a training method of an exemplary embodiment of the present disclosure is shown. As shown in the flowchart, the training method of the exemplary embodiment of the present disclosure can include: Figure 1

[0057] Step 101: obtaining graph structure sample data based on user basic information and heterogeneous behavior information of a user for a multimedia product, the graph structure sample data including node data of a plurality of node samples, a plurality of behavior sample data, and relationship information of the node data, the relationship information of the node data including a heterogeneous behavior relationship between a user sample and a multimedia product sample. Based on this, the present disclosure can introduce high-order heterogeneous behavior of a user for a multimedia product into the graph structure sample data to form a unified heterogeneous graph with a user and a multimedia product as nodes and a user behavior as an edge.

[0058] In actual application, a large amount of user basic information and heterogeneous behavior of each user for a multimedia product in a multimedia playing platform can be collected, and then a user sample set U is constructed based on a large amount of user basic information, and a multimedia product sample set V and a behavior sample set R are constructed based on the heterogeneous behavior of each user for a multimedia product.

[0059] The user sample set U includes attribute information of a plurality of user samples u, and the capacity of the user sample set U can be represented by |U|. At this time, the node data of the user sample can include the attribute information of the user sample. These user basic information can be user attribute information, such as user identity information such as user age, gender, and education, and user terminal information such as a type of terminal held by the user and positioning information.

[0060] ​The multimedia product sample set V includes attribute information of multiple multimedia product samples v, and the capacity of the multimedia product sample set V can be represented by |V|. At this time, the node data of the multimedia product sample is the attribute information of the multimedia product sample. The multimedia product of the exemplary embodiments of the present disclosure can be a video product, an audio product, a graphic-text product, or the like, but is not limited thereto. Taking a video product as an example, the multimedia product sample set V includes identity information of multiple video products, and the identity information can be divided into film and television, variety show, science and education film, documentary, animation, sports, and the like according to the fields involved in the content of the video products.

[0061] The behavior sample set R includes identity information of multiple behavior samples r, and at this time, the behavior sample data can be attribute information of the behavior sample, which acts as the role of the edge sample data in the graph structure sample data. Taking a video product as an example, considering that a user in a video playing platform can exert heterogeneous behaviors (i.e., multiple behaviors) on the same video product, the identity information of various behavior samples can be obtained from the heterogeneous behavior information of the multimedia product collected from the video playing platform, which can include multiple behaviors such as browsing and staying, jump source, clicking, playing time, collecting, commenting, completing playing, downloading, and forwarding.

[0062] After the user sample set U, the multimedia product sample set V, and the behavior sample set R are extracted, the attribute information of the user sample u included in the user sample set U, the attribute information of the multimedia product sample v, and the attribute information of the behavior sample r can be obtained based on the heterogeneous behavior information of the user on the multimedia product, and the relationships between the user sample u, the multimedia product sample v, and the behavior sample r are determined, and these relationships are defined as a user-item heterogeneous interaction set, which can be represented by {Y (1) ,Y (2) ,…,Y (K)} where K represents the total number of types of behavior samples, Y (k) =[y(k)uv], Y (1) represents the first behavior interaction matrix of the user-item, Y (2) represents the second behavior interaction matrix of the user-item, and Y (k) represents the kth behavior interaction matrix of the user-item.

[0063] y(k)uv is an element in the matrix Y (k) , which represents whether the user sample u has the kth behavior sample with the multimedia product sample v, and the capacity of y(k)uv is |U|×|V|, y(k)uv∈{0,1}. When y(k)uv=1, it indicates that the user sample u has the kth behavior sample with the multimedia product sample v, and when y(k)uv=1, it indicates that the user sample u does not have the kth behavior sample with the multimedia product sample v.

[0064] Step 102: training the behavior prediction model based on the graph structure sample data, the behavior prediction model comprising a graph neural network and a behavior preference prediction layer, the graph neural network being configured to obtain node update data of a node sample based on node data and behavior sample data of adjacent node samples associated with the same node sample, and the behavior preference prediction layer being configured to determine multi-behavior preference prediction information of a plurality of user samples for different multimedia product samples based on a plurality of behavior sample data and node update data of a plurality of node samples.

[0065] When the current node sample is a user sample, the neighbor node can be a multimedia product sample. At this time, the graph neural network is configured to determine node update data of the corresponding user sample based on node data and behavior sample data of the multimedia product sample associated with the same user sample. Here, the multimedia product sample associated with the user sample can be a multimedia product connected by an edge, which can include one or more multimedia products, and the behavior sample associated with the user sample can be a behavior sample associated with the user sample as an edge, which can include a plurality of behavior samples.

[0066] When the current node sample is a multimedia product sample, the neighbor node sample is a user sample. At this time, the graph neural network is configured to determine node update data of the corresponding multimedia product sample based on node data and behavior sample data of the user sample associated with the same multimedia product sample. Here, the user sample associated with the multimedia product sample can be a user sample that has a behavior with the multimedia product sample, and the behavior sample data associated with the multimedia product sample can be a behavior applied to the multimedia product sample.

[0067] Considering that there is a heterogeneous behavior relationship between each user sample and a plurality of multimedia product samples, that is, multi-behavior, there is a high-order connection between the user sample and the multimedia product sample bridged by multi-behavior, therefore, the behavior prediction model can not only fuse heterogeneous behavior information into users and multimedia products, but also consider the high-order connection between users, multimedia products and user behaviors through the information transmission process, in order to fully analyze the sparse behavior information of users on target type multimedia products, thereby effectively extracting multi-behavior preference prediction information of users for target type multimedia products, so as to achieve the purpose of deep modeling of high-order heterogeneous behavior.

[0068] The graph neural network is used for determining node update data of each node sample based on node data of neighbor node samples adjacent to each node sample in an information transmission process, and does not fuse the node data of the node sample, so as to reduce the model parameter quantity, make the behavior prediction model lightweight, and improve the behavior preference prediction speed of the behavior prediction model. Meanwhile, the self-connection of the node sample is captured by means of the interlayer connection operation of the neural network, so that the node data loss problem caused by fusing only the node data of the neighbor node sample can be avoided. Therefore, the behavior preference prediction layer in the example embodiment of the disclosure can accurately determine the multiple behavior preference prediction information of the user sample for the multimedia product sample based on the multiple behavior sample information and the node update information of the multiple node samples, so as to make sufficient preparation for multimedia product recommendation, ensure accurate recommendation of the multimedia product matched with the target user, and improve the user retention rate of the multimedia product playing platform.

[0069] In an optional manner, the graph neural network in the example embodiment of the disclosure is used for determining node composite data of each node sample based on node data and behavior sample data of neighbor node samples adjacent to each node sample, performing graph convolution operation on the node composite data of the multiple node samples, and obtaining node update data of the multiple node samples.

[0070] In actual application, for a current node sample, neighbor node samples adjacent to the current node sample can be obtained from the node data of the multiple node samples based on the relationship information of the node data. The neighbor node sample here can refer to a node sample directly connected to the current node sample in a graph structure. Similarly, behavior sample data related to the current node sample can be obtained from the multiple behavior sample data based on the relationship information of the node data. The behavior sample data here can refer to behavior sample data acting as a relationship bridge between the current node sample and the adjacent node sample.

[0071] The data processing flow of the graph convolutional neural network in the example embodiment of the disclosure can be divided into a composite phase and a graph convolution phase. In the composite phase, the node data and the behavior sample data of the neighbor node sample are multiplied to achieve the purpose of combining the node data and the behavior sample data of the neighbor node sample. In the graph convolution phase, since the graph convolutional neural network can fuse the node composite data of each node sample and the node composite data of the neighbor node sample, in the composite phase, the node data of each node sample only fuses the node data and the behavior sample data of the neighbor node sample, but in the graph convolution phase, the node data of each node sample can be captured by fusing the node composite data of each node sample and the node composite data of the neighbor node sample, so as to ensure that the finally obtained node update data of the node sample is complete and reliable.

[0072] Exemplarily, when the graph neural network of the exemplary embodiment of the present disclosure performs the graph convolution operation on the node composite data of the plurality of node samples, the node composite data of each node sample can be aggregated based on the normalized adjacency matrix to obtain the node aggregated data of the corresponding node sample, and the node aggregated data of the plurality of node samples can be fused by the neural network to obtain the node updated data of the plurality of node samples.

[0073] The normalized adjacency matrix can be determined based on the relationship information of the node data. For example, the adjacency matrix A and the degree matrix D can be obtained based on the relationship information of the node data, and then the normalized adjacency matrix can be obtained based on the degree matrix and the adjacency matrix. The normalized adjacency matrix can be determined based on the relationship information of the node data. For example, the adjacency matrix A and the degree matrix D can be obtained based on the relationship information of the node data, and then the normalized adjacency matrix can be obtained based on the degree matrix and the adjacency matrix. The normalized adjacency matrix can be determined based on the relationship information of the node data. For example, the adjacency matrix A and the degree matrix D can be obtained based on the relationship information of the node data, and then the normalized adjacency matrix can be obtained based on the degree matrix and the adjacency matrix.

[0074]

[0075] Wherein, I is the unit matrix.

[0076] At this time, the node data of the node sample and the node data of the adjacent node sample can be aggregated based on the normalized adjacency matrix and the node composite data of the node sample to obtain the node aggregated data of the node sample. In this fusion process, the node data of the node sample that is not compounded in the compounding stage can be captured by the normalized adjacency matrix The node composite data of each node sample and the composite data of the neighbor node are aggregated, so that each node sample captures the node data of the node sample that is not compounded in the compounding stage, and the obtained node aggregated data of the node sample is complete and reliable.

[0077] The neural network of the exemplary embodiment of the present disclosure can include a plurality of fully connected layers, and the node aggregated data of the plurality of node samples can be fused by the high receptive field of the plurality of fully connected layers. This not only allows each node sample to capture the node data of the node sample that is not compounded in the compounding stage, but also ensures that the information amount of the obtained node updated data of each node sample is more abundant.

[0078] The graph neural network of the exemplary embodiment of the present disclosure can be a multi-layer graph neural network, and the node data of the plurality of node samples input into each layer of the graph neural network can be the node updated data of the plurality of node samples output from the previous layer of the graph neural network. For example, the node data of the user sample input into the current layer of the graph neural network can be the node updated data of the user sample input into the previous layer of the graph neural network, and the node data of the multimedia product sample input into the current layer of the graph neural network can be the node updated data of the multimedia product input into the previous layer of the graph neural network.

[0079] In order to make the graph neural network output the node update data of the user sample, the node update data of the multimedia product sample and the behavior sample update data with the same embedding space dimension, the graph neural network of the example embodiment of the present disclosure is further used to project the behavior sample data into the embedding space of the node update data of the user sample and the node update data of the multimedia product sample, to obtain the behavior sample update data. When the graph neural network is a multi-layer graph neural network, the behavior sample data input into each layer of the graph neural network is the behavior sample update data output by the previous layer of the graph neural network.

[0080] For example, the embedding space of the behavior sample data is projected into the embedding space dimension of the node update data of the user sample and the node update data of the multimedia product sample, so as to ensure that the graph neural network can output the node update data of the user sample, the node update data of the multimedia product sample and the behavior sample update data with the same embedding space dimension.

[0081] For example, the example embodiment of the present disclosure can introduce layer-specific parameters in each layer of the graph neural network, and the layer-specific parameters can convert the embedding space of the input behavior sample data and project it into the embedding space of the node update data of the user sample and the node update data of the multimedia product sample.

[0082] The determination process of the node update data output by the lth layer of the graph neural network is described below.

[0083] The node update data of the user sample output by the lth layer of the graph neural network is represented as formula two:

[0084]

[0085] The node update data of the multimedia product sample output by the lth layer of the graph neural network is represented as formula three:

[0086]

[0087] In formula two and formula three, W (l) represents the model parameters of the lth layer of the graph neural network, and the normalized adjacency matrix is used to limit the embedding scale under the increase of the graph convolution operation, and |N u represents the number of neighbors of the user sample u in the graph structure sample data, and |N v represents the number of neighbors of the multimedia product sample v in the graph structure sample data, and φ represents a composite operation function, and the composite method can be a dot product operation.

[0088] In formula two, the composite operation function φ can convert the behavior sample update data output by the (l-1)th layer of the graph neural network The node update data of the multimedia product sample output by the (l-1)th layer graph neural network The composite operation function φ in formula three can update the behavior sample update data output by the (l-1)th layer graph neural network The node update data of the user sample output by the (l-1)th layer graph neural network The composite operation function φ in formula three can update the behavior sample update data output by the (l-1)th layer graph neural network

[0089] The node update data of the multimedia product sample output by the (l-1)th layer graph neural network The formula three is as follows:

[0090]

[0091] The layer-specific parameter of the (l)th layer graph neural network, which can be a constant parameter or a to-be-updated parameter that needs to be updated by training the behavior prediction model.

[0092] In an optional manner, when the graph neural network of the example embodiment of the present disclosure is a multi-layer graph neural network, the behavior preference prediction layer can be used to determine the node target data of the corresponding node sample based on the node update data of the same node sample output by the multi-layer graph neural network, and determine the multiple behavior preference prediction information of the multiple user samples for different multimedia product samples based on the node target data of the multiple node samples and the multiple behavior sample data.

[0093] In actual applications, when the example embodiment of the present disclosure determines the node target data of the corresponding node sample based on the node update data of the same node sample output by the multi-layer graph neural network, each node target data can be determined by the weighted sum result of the node update data of the corresponding node sample output by the multi-layer graph neural network. It can be seen that the example embodiment of the present disclosure can fuse the node update data of the same node sample output by the graph neural networks of different layers, ensure that the information amount of the node target data of the node sample is relatively rich and complete, and thus improve the prediction accuracy of the subsequent behavior preference prediction information.

[0094] When the node update data of the corresponding node sample output by the multi-layer graph neural network is weighted and summed, the node update data weight of the corresponding sample structure output by each layer of graph neural network can be the same. Taking the number of layers of the graph neural network as L layers as an example, the node target data e u of the user sample is described by formula five, the node target data e v of the multimedia product sample is described by formula six, and the behavior sample target data e r .

[0095]

[0096]

[0097]

[0098] when l = 0, denotes the node data of the user sample, denotes the node data of the multimedia product sample, denotes the behavior sample data, when l ≥ 1, denotes the node update data of the user sample output by the l-th layer graph neural network, denotes the node update data of the multimedia product sample output by the l-th layer graph neural network, denotes the behavior sample update data output by the l-th layer graph neural network.

[0099] when the node target data e u of the user sample, the node target data e v of the multimedia product sample, and the behavior sample target data e r of the example embodiment of the present disclosure are represented in vector form, the behavior preference prediction information of the user sample u for the k-th behavior sample of the multimedia product sample v can be determined by Equation Eight

[0100]

[0101] i denotes the element position of the vector, d denotes the dimension of the vector, denotes the transposition of the vector representation of the node target data e u of the user sample u, denotes the vector representation of the k-th behavior sample target data e v denotes the node target data of the multimedia product sample v, which is represented in vector form in Equation Eight e u,i denotes the i-th element in the vector representation of the node target data of the user sample u, denotes the i-th element in the vector representation of the k-th behavior sample target data e v,i denotes the i-th element in the vector representation of the node target data of the multimedia product sample v.

[0102] In an optional manner, the learning manner of the behavior prediction model is a negative sampling free learning manner when the example embodiment of the present disclosure trains the behavior prediction model. The example embodiment of the present disclosure trains the behavior prediction model based on the graph structure sample data, including:

[0103] The graph neural network obtains node update data of the node sample based on node data and behavior sample data of adjacent node samples associated with the same node sample. The behavior preference prediction layer determines multiple behavior preference prediction information of the user sample for the multimedia product sample based on multiple behavior sample data, node update data of multiple user samples, and node update data of multiple multimedia product samples. Multiple behavior sample preference labels of the user sample for different multimedia product samples are extracted from relationship information of the node data. The loss of the behavior prediction model is determined based on the multiple behavior sample preference labels of the user sample for different multimedia product samples and the multiple behavior preference prediction information. The behavior prediction model is updated based on the loss of the behavior prediction model.

[0104] In actual application, the user sample performs a certain behavior sample on the multimedia product, and it can be considered that the behavior sample label of the user sample for the multimedia product sample is 1, otherwise it is 0. Based on this, the behavior sample preference label of each user sample for the multimedia product sample can be extracted from the relationship information of the node data. Considering that the learning mode of the behavior prediction model of the example embodiment of the present disclosure is a multi-task learning mode, the loss of the behavior prediction model is determined by the loss of multiple behavior samples. For example, the loss of the kth behavior sample can be determined by formula ten

[0105]

[0106] In formula ten, y(k) represents the kth behavior sample importance parameter of the user sample u for the multimedia product sample v, uv y(k) represents the kth behavior sample preference label of the user sample u for the multimedia product sample v, y(k) represents the kth behavior sample preference information of the user sample u for the multimedia product sample v.

[0107] The determination method of the loss of each behavior sample can include: extracting multiple behavior sample preference labels of the user sample for the multimedia product sample based on the relationship information of the node data; determining the behavior sample sub-loss of each user sample for the corresponding multimedia product sample based on the behavior preference prediction information and the behavior sample preference label of each user sample for the same multimedia product sample; and determining the loss of the corresponding behavior sample based on the behavior sub-loss of each user sample for the same behavior sample of different multimedia product samples.

[0108] For example, when determining the behavior sample sub-loss of each user sample for the corresponding multimedia product sample based on the behavior preference prediction information and the behavior sample preference label of each user sample for the same multimedia product sample, the behavior deviation of the corresponding user sample for the same multimedia product sample is determined based on the behavior preference prediction information and the behavior sample preference label of each user sample for the same multimedia product sample, and the loss of the behavior sample of each user sample for the corresponding multimedia product sample is determined based on the behavior deviation of each user sample for the same multimedia product sample and the behavior sample importance parameter.

[0109] The method of the example embodiment of the present disclosure can obtain the known multiple behavior relationships between the user samples and the multimedia product samples from the relationship information of the node data, and then determine the multiple behavior sample preference labels of each user sample for the multimedia product samples. The behavior sample importance parameter can be the behavior importance of the user sample for the corresponding multimedia product sample, for example, the importance of the comment behavior is greater than the importance of the like behavior, and the importance of the like behavior is greater than the importance of the forward behavior. Meanwhile, the same behavior sample importance parameter of the same user sample for different multimedia product samples can be the same or different, and the same behavior sample importance parameter of different samples for the same multimedia product sample can be the same or different.

[0110] For example, the loss L(θ) of the behavior prediction model of the example embodiment of the present disclosure can be determined by using Formula Nine:

[0111]

[0112] In Formula Nine, λ k represents the first weight, μ represents the second weight, represents the loss of the kth behavior sample, θ represents the model parameter of the behavior prediction model, and the introduction of the regularization can reduce the probability of overfitting.

[0113] After determining the loss of the behavior prediction model, it can be judged whether the loss of the behavior prediction model is less than the loss threshold. If it is less than the loss threshold, it can be considered that the training is completed, otherwise, the model parameter of the behavior prediction model in the multi-task learning mode can be optimized by the Adam optimizer. The model parameter here can be the model parameter of the graph convolution network.

[0114] Figure 2 A flowchart of a recommendation method of the example embodiment of the present disclosure is shown. As Figure 2 shown, the recommendation method provided by the example embodiment of the present disclosure can include:

[0115] Step 201: input the target user data, the target behavior data and the plurality of multimedia product data into the behavior prediction model of the exemplary embodiment of the present disclosure.

[0116] In actual application, the various information input into the behavior prediction model can be collected from the multimedia product playing platform, the target user data and the multimedia product data can be collected through the multimedia product playing platform, the target user data can be the information of the user who needs to be recommended the multimedia product, which can include the gender, age and the like of the target user. The multimedia product data can be a plurality of multimedia product data of the same category, for example, for video products, a variety of video products can be collected, including but not limited to video products with a daily browsing frequency greater than or equal to a preset threshold and video products with a daily browsing frequency less than the preset threshold.

[0117] The target behavior data can be a preset behavior information, which can come from the multimedia playing platform, for example, the category of the target behavior can reflect that the target user pays more attention to the multimedia product, which can be determined by the behavior sample importance parameter, for example, one or more behavior data with a sample importance greater than a certain threshold determined by the sample importance parameter can be selected as the target behavior data.

[0118] Step 202: obtain the multiple behavior preference prediction information of the target user for different multimedia product samples from the multiple user samples for different multimedia product samples by using the behavior prediction model.

[0119] When the behavior prediction model training is completed, it can determine the multiple behavior preference prediction information of the plurality of user samples for different multimedia product samples under the optimal model parameters, that is, the high-order connection among the user samples, the multimedia product samples and the user behavior samples. Considering that the various information input into the behavior prediction model can be collected from the multimedia product playing platform, the target user data, the multimedia product data and the target behavior data also come from the multimedia product playing platform, therefore, the multiple behavior preference prediction information of the target user for different multimedia product samples can be found by using the behavior prediction model to obtain the multiple behavior preference prediction information of the plurality of user samples for different multimedia product samples.

[0120] Step 203: if the target behavior preference information of the target user for the multimedia product meets the recommendation condition, recommend the target multimedia product.

[0121] In actual applications, the target behavior data of the example embodiments of the present disclosure can include one target behavior data or multiple target behavior data. When multiple target behavior data are included, the target behavior preference index of the target user for the target multimedia product includes a weighted result of multiple behavior prediction probabilities of the target user for the multimedia product. At this time, the multiple target behavior prediction probabilities can be introduced into the recommended multimedia product to ensure the recommendation accuracy.

[0122] In an embodiment, the target behavior preference index of the target user for the multimedia product includes a target behavior prediction probability of the target user for the multimedia product. At this time, the recommendation condition of the example embodiments of the present disclosure can include at least one of the following two conditions:

[0123] The first condition is that the target behavior prediction probability of the target user for the multimedia product is greater than or equal to a preset probability. That is, as long as the target behavior prediction probability of the target user for a certain multimedia product is greater than or equal to a preset probability, it is considered that the interest of the target user for the multimedia product is relatively high. The multimedia product category can be determined based on the target data of the multimedia product, and the multimedia product can be recommended based on the multimedia product category.

[0124] The second condition is that the multimedia products are sorted in descending order of the target behavior prediction probability of the target user for the multimedia products, and the target multimedia product includes the first f multimedia products.

[0125] In one or more technical solutions provided in the example embodiments of the present disclosure, the graph structure sample data can be obtained based on the user basic information and the heterogeneous behavior information of the user for the multimedia product, so that the graph structure sample data includes node data of multiple node samples, multiple behavior sample data, and relationship information of the node data. It can be seen that the high-order heterogeneous behavior of the user for the multimedia product can be introduced into the graph structure sample data to form a heterogeneous graph with the user sample and the multimedia product sample as nodes and the heterogeneous behavior between the user sample and the multimedia product sample as edges.

[0126] When the behavior prediction model is trained based on the multi-relation graph structure data, the graph neural network is used to determine the node update data of the corresponding target node sample based on the node data of the neighbor node sample adjacent to each target node sample, and the behavior preference prediction layer is used to determine the multiple behavior preference prediction information of the user sample for the multimedia product sample based on the multiple behavior sample data and the node update data of the multiple node samples. It can be seen that the behavior prediction model can not only fuse heterogeneous behavior information into the user and the multimedia product, but also consider the high-order connection between the user, the multimedia product and the user behavior through the information transmission process, so as to fully analyze the sparse behavior information of the user for the target type multimedia product, effectively extract the multiple behavior preference prediction information of the user for the target type multimedia product, and achieve the purpose of deep modeling of high-order heterogeneous behavior.

[0127] Furthermore, the graph neural network is used to determine the node update data of the corresponding node sample based on the node data of the neighbor node sample adjacent to each node sample, and the node data of the node sample is not fused to reduce the amount of model parameters, so that the behavior prediction model is lightweight and the behavior preference prediction speed of the behavior prediction model is improved. At the same time, the self-connection of the node sample is captured through the interlayer connection operation of the graph neural network, which can ensure that the node sample will not cause node data loss due to only fusing the node data of the neighbor node sample. Therefore, the behavior preference prediction layer of the example embodiment of the disclosure can accurately determine the multiple behavior preference prediction information of the user sample for the multimedia product sample based on the multiple behavior sample information and the node update information of the multiple node samples, and fully prepare for multimedia product recommendation.

[0128] The method of the example embodiment of the disclosure can determine the initial node data of the target user, the initial node data of the multimedia product and the initial behavior node data based on the heterogeneous behavior information of the target user for the multiple multimedia products in the multimedia product recommendation scene, and then input the initial node data of the target user, the initial node data of the multimedia product and the initial behavior node data into the trained behavior prediction model, so as to obtain the multiple behavior preference information of the target user for the multiple multimedia products. Then, it is judged whether the target behavior preference information of the target user for the target multimedia product meets the recommendation condition based on the collaborative filtering mode. If the recommendation condition is met, it means that the target user has a high possibility to apply the target behavior to the target multimedia product, and the target multimedia product can be recommended.

[0129] It can be seen that in the method of the example embodiment of the present disclosure, if the target user's behavior information for a certain type of multimedia product is sparse due to personal daily preference problems, the target user data, target behavior data and multiple multimedia product data can be input into the trained behavior prediction model to use the behavior prediction model to predict multiple behavior preferences from multiple user samples for different multimedia product samples, that is, high-order connections between user samples, multimedia product samples and behavior samples, analyze the sparse behavior information of the target user for the multimedia product, and obtain the multiple behavior preference information of the target user for multiple multimedia products. Then, based on this, the target multimedia product with sparse behavior information is accurately recommended to the target user, thereby increasing the attractiveness of the recommended target multimedia data to the target user and improving the user retention rate of the multimedia playing platform.

[0130] The above mainly introduces the scheme provided by the example embodiment of the present disclosure from the perspective of the electronic device. It can be understood that the electronic device contains a hardware structure and / or a software module corresponding to the execution of each function in order to realize the above functions. Those skilled in the art should easily realize that the units and algorithm steps of each example described in combination with the embodiments disclosed in the present text can be realized in the form of hardware or a combination of hardware and computer software. Whether a certain function is realized in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present disclosure.

[0131] The example embodiment of the present disclosure can divide the functional units of the electronic device according to the above method, for example, each functional module can be divided according to each function, or two or more functions can be integrated in one processing module. The integrated module can be realized in the form of hardware or in the form of a software functional module. It should be noted that the division of the module in the example embodiment of the present disclosure is illustrative, and is only a logical functional division. When actually implemented, there can be another division method.

[0132] In the case of dividing each functional module according to each function, the example embodiment of the present disclosure provides a training device, which can be an electronic device or a chip applied to an electronic device. Figure 3 A functional module schematic block diagram of the training device according to the example embodiment of the present disclosure is shown. As shown in Figure 3 The training device 300 includes:

[0133] The acquisition module 301 is configured to acquire graph structure sample data based on the user basic information and the heterogeneous behavior information of the user for the multimedia product, the graph structure sample data including node data of a plurality of node samples, a plurality of behavior sample data, and relationship information of the node data, the plurality of node samples including a user sample and a multimedia product sample, the behavior sample data including a plurality of, and the relationship information of the node data including a heterogeneous behavior relationship between the user sample and the multimedia product sample.

[0134] The training module 302 is configured to train a behavior prediction model based on the graph structure sample data, the behavior prediction model including a graph neural network and a behavior preference prediction layer, the graph neural network being configured to obtain node update data of a node sample based on node data and behavior sample data of adjacent node samples associated with the node sample, and the behavior preference prediction layer being configured to determine multiple behavior preference prediction information of multiple user samples for different multimedia product samples based on multiple behavior sample data, node update data of the multiple user samples, and node update data of the multiple multimedia product samples.

[0135] In an optional manner, the graph neural network is configured to obtain node composite data corresponding to the node sample based on the node data and the behavior sample data of the adjacent node samples associated with the node sample, perform graph convolution operation on the node composite data of the multiple node samples, and obtain node update data of the multiple node samples.

[0136] In an optional manner, the graph neural network is configured to aggregate the node composite data of each node sample based on a normalized adjacency matrix to obtain node aggregation data corresponding to the node sample, and fuse the node aggregation data of the multiple node samples through a neural network to obtain node update data of the multiple node samples, the normalized adjacency matrix being determined based on the relationship information of the node data.

[0137] In an optional manner, the graph neural network is configured to obtain node update data corresponding to the user sample based on the node data and the behavior sample data of the multimedia product sample associated with the user sample, and determine node update data corresponding to the multimedia product sample based on the node data and the behavior sample data of the user sample associated with the multimedia product sample.

[0138] In an optional manner, the graph neural network is a multi-layer graph neural network, the behavior preference prediction layer is configured to determine node target data of the node sample based on node update data of the node sample output by the multi-layer graph neural network, and determine multiple behavior preference prediction information of the user sample for different multimedia product samples based on the node target data of the multiple node samples and the multiple behavior sample data.

[0139] In an optional manner, each node target data is determined by weighted summation of node update data of the corresponding node sample output by the multi-layer graph neural network.

[0140] In an optional manner, the learning manner of the behavior prediction model is a negative sampling free learning manner, the loss of the behavior prediction model is determined by the loss of the multiple behavior samples, and the training module 301 is configured to:

[0141] The graph neural network obtains node update data of the node sample based on node data and behavior sample data of adjacent node samples associated with the same node sample, the behavior preference prediction layer determines multiple behavior preference prediction information of the user sample for a multimedia product sample based on the multiple behavior sample data, the node update data of the multiple user samples, and the node update data of the multiple multimedia product samples, extracts multiple behavior sample preference labels of the multiple user samples for different multimedia product samples from the relationship information of the node data, and determines the loss of the behavior prediction model based on the multiple behavior sample preference labels of the multiple user samples for different multimedia product samples and the multiple behavior preference prediction information.

[0142] The behavior prediction model is updated based on the loss of the graph neural network.

[0143] In the case of dividing each functional module according to each function, the exemplary embodiments of the present disclosure provide a recommendation device, which can be a server or a chip applied to a server. Figure 4 A functional module schematic block diagram of a recommendation device according to an exemplary embodiment of the present disclosure is shown. As shown in Figure 4 The recommendation device 400 includes:

[0144] An inference module 401 is configured to input target user data, target behavior data, and multiple multimedia product data into the behavior prediction model trained by the exemplary embodiments of the present disclosure, and obtain multiple behavior preference prediction information of the target user for different multimedia product samples from the behavior prediction model based on multiple behavior preference prediction information of the multiple user samples for different multimedia product samples.

[0145] The recommendation module 402 is configured to recommend the target multimedia product if the target behavior preference information of the target user for the target multimedia product meets a recommendation condition.

[0146] In an optional manner, the target behavior preference information of the target user for the multimedia product includes a target behavior probability of the target user for the multimedia product, and the recommendation condition includes:

[0147] The target behavior prediction probability of the target user for the multimedia product is greater than or equal to a preset probability; and / or,

[0148] The multimedia products are sorted according to the target behavior prediction probability of the target user for the multimedia product in descending order, and the target multimedia product includes the first f multimedia products, where f represents an integer greater than or equal to 1 and less than the total number of multimedia products.

[0149] In an optional manner, the target behavior data includes a plurality of target behavior data, and the target behavior preference index of the target user for the target multimedia product includes a weighted result of a plurality of target behavior prediction probabilities of the target user for the multimedia product.

[0150] Figure 5 A schematic block diagram of a chip according to an example embodiment of the present disclosure is shown. As shown in the figure, the chip 500 includes one or more than two (including two) processors 501 and a communication interface 502. The communication interface 502 can support the server to perform the data transceiving steps in the above method, and the processor 501 can support the server to perform the data processing steps in the above method. Figure 5

[0151] Optionally, as shown in the figure, the chip 500 further includes a memory 503, which can include a read-only memory and a random access memory, and provides operation instructions and data for the processor. Part of the memory can also include a non-volatile random access memory (NVRAM). Figure 5

[0152] In some embodiments, as shown in the figure, the chip 500 further includes a bus architecture, which includes a processor bus, a peripheral bus, and a graphics bus. Figure 5 ​​As shown, the processor 501 executes various steps of the methods disclosed in embodiments of the present disclosure by calling stored instructions. The processor 501 controls the processing operations of any of the terminal devices, and can also be referred to as a central processing unit (CPU). The memory 503 can include a read-only memory and a random access memory, and provides instructions and data to the processor 501. A portion of the memory 503 can also include a NVRAM. The memory, the communication interface, and the bus system are coupled together via a bus system, which can include a data bus, a power supply bus, a control bus, and a state signal bus, etc. However, for the sake of clarity, only the data bus is shown in Figure 5 FIG. 4.

[0153] The method disclosed in the embodiments of the present disclosure can be applied to a processor, or implemented by a processor. The processor can be an integrated circuit chip having a signal processing capability. In the implementation process, the steps of the above method can be completed by hardware integrated logic circuit or software form of instructions in the processor. The processor mentioned above can be a general processor, a digital signal processor (DSP), an ASIC, a field-programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component. The disclosed methods, steps and logic block diagrams in the embodiments of the present disclosure can be implemented or executed. The general processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in conjunction with the embodiments of the present disclosure can be directly embodied as a hardware code processor for execution, or a combination of hardware and software modules in the code processor for execution. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, or other mature storage media in the art. The storage medium is located in the memory, and the processor reads the information in the memory and combines the hardware to complete the steps of the above method.

[0154] The exemplary embodiments of the present disclosure also provide an electronic device, including at least one processor, and a memory connected with the at least one processor in communication. The memory stores a computer program capable of being executed by the at least one processor, and the computer program, when executed by the at least one processor, is configured to cause the electronic device to perform the method according to the embodiments of the present disclosure.

[0155] The exemplary embodiments of the present disclosure further provide a non-transitory computer readable storage medium storing a computer program, wherein the computer program, when executed by a processor of a computer, causes the computer to perform the method according to the embodiments of the present disclosure.

[0156] The exemplary embodiments of the present disclosure further provide a computer program product comprising a computer program, wherein the computer program, when executed by a processor of a computer, causes the computer to perform the method according to the embodiments of the present disclosure.

[0157] Reference Figure 6 The structure block diagram of the electronic device 600, which can be applied to various aspects of the present disclosure, will now be described, which is an example of a hardware device that can be applied to the present disclosure. The electronic device is intended to represent a variety of forms of digital electronic computing devices such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computing devices. The electronic device can also represent a variety of forms of mobile devices such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are meant to be examples only, and are not intended to limit the implementations of the present disclosure described and / or claimed in this document.

[0158] As shown in Figure 6 The electronic device 600 includes a computing unit 601 that can perform various appropriate actions and processes in accordance with a computer program stored in a read-only memory (ROM) 602 or a computer program loaded from a storage unit 608 into a random access memory (RAM) 603. In the RAM 603, various programs and data required for the operation of the device 600 can also be stored. The computing unit 601, the ROM 602, and the RAM 603 are connected to each other through a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0159] As shown in Figure 6As shown, multiple components in electronic device 600 are connected to I / O interface 605, including: input unit 606, output unit 607, storage unit 608, and communication unit 609. Input unit 606 can be any type of device capable of inputting information to electronic device 600. Input unit 606 can receive input digital or character information and generate key signal inputs related to user settings and / or function control of electronic device. Output unit 607 can be any type of device capable of presenting information and may include, but is not limited to, a display, speaker, video / audio output terminal, vibrator, and / or printer. Storage unit 608 may include, but is not limited to, disk and optical disk. Communication unit 609 allows electronic device 600 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks, and may include, but is not limited to, modems, network cards, infrared communication devices, wireless communication transceivers and / or chipsets, such as Bluetooth™ devices, WiFi devices, WiMax devices, cellular communication devices, and / or the like.

[0160] like Figure 6 As shown, computing unit 601 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of computing unit 601 include, but are not limited to, central processing unit (CPU), graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Computing unit 601 performs the various methods and processes described above. For example, in some embodiments, the methods of exemplary embodiments of this disclosure can be implemented as computer software programs tangibly contained in a machine-readable medium, such as storage unit 608. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 600 via ROM 602 and / or communication unit 609. In some embodiments, computing unit 601 can be configured to perform the methods of exemplary embodiments of this disclosure by any other suitable means (e.g., by means of firmware).

[0161] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0162] In the context of this disclosure, a machine-readable medium can be a tangible medium that contains or stores a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include but is not limited to an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of the machine-readable storage medium will include one or more lines of electrical connections, portable computer disks, hard disk drives, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or Flash memory), optical fibers, portable compact disc read-only memories (CD-ROMs), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0163] As used in this disclosure, the terms "machine-readable medium" and "computer- readable medium" refer to any computer program product, apparatus and / or device (e.g., magnetic discs, optical disks, memory, Programmable Logic Devices (PLDs)) used to provide machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal that can be used to provide machine instructions and / or data to a programmable processor.

[0164] To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0165] The systems and techniques described here can be implemented in a computing system that includes a back end component, e.g., as a data server, or that includes a middleware component, e.g., an application server, or that includes a front end component, e.g., a user computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here, or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication, e.g., a communication network. Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0166] The computer system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.

[0167] In the embodiments described above, the whole or part of the embodiments can be realized by software, hardware, firmware, or any combination thereof. When realized by software, the whole or part of the embodiments can be realized in the form of a computer program product. The computer program product includes one or more computer programs or instructions. When the computer programs or instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present disclosure are performed wholly or partially. The computer can be a general purpose computer, a special purpose computer, a computer network, a terminal, user equipment, or other programmable apparatus. The computer programs or instructions can be stored in a computer readable storage medium or transferred from one computer readable storage medium to another computer readable storage medium, for example, the computer programs or instructions can be transferred from one website site, computer, server, or data center to another website site, computer, server, or data center through a wired or wireless manner. The computer readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, and the like that integrates one or more available media. The available media can be a magnetic medium, such as a floppy disk, a hard disk, a magnetic tape; an optical medium, such as a digital video disc (DVD); and a semiconductor medium, such as a solid state disk (SSD).

[0168] Although the present disclosure has been described in connection with certain specific features and embodiments thereof, it is to be understood that it is provided as an exemplification of the principles of the present disclosure and the features set forth herein are intended to be illustrative rather than limiting, and that numerous modifications and variations therein can be expected by those skilled in the art. Accordingly, it should be understood that the description and drawings are illustrative of the present disclosure and are not intended to be limiting. It should be understood that various changes can be made to the implementations described and the embodiments presented herein without departing from the spirit and scope of the present disclosure. It is intended that all such changes be considered as within the scope of the present disclosure.

Claims

1. A training method, characterized in that, include: Graph structure sample data is obtained based on user basic information and heterogeneous user behavior information for multimedia products. The graph structure sample data includes node data of multiple node samples, multiple behavior sample data and relationship information of node data. The multiple node samples include user samples and multimedia product samples. The relationship information of node data includes heterogeneous behavior relationships between user samples and multimedia product samples. A behavior prediction model is trained based on the graph structure sample data. The behavior prediction model includes a graph neural network and a behavior preference prediction layer. The graph neural network is used to obtain node update data of the node sample based on the node data and behavior sample data of the neighboring node samples associated with the same node sample. The behavior preference prediction layer is used to determine multiple behavior preference prediction information of multiple user samples for different multimedia product samples based on multiple behavior sample data, multiple user sample node update data, and multiple multimedia product sample node update data. The graph neural network is used to obtain node update data corresponding to the user sample based on node data and behavior sample data of the multimedia product sample associated with the same user sample, and to determine node update data corresponding to the multimedia product sample based on node data and behavior sample data of the user sample associated with the same multimedia product sample. The data processing flow of the graph neural network is divided into a composite stage and a graph convolution stage. In the composite stage, the node data of each node sample is fused with the behavioral sample data associated with each node sample. In the graph convolution stage, the node data of each node sample is fused with the composite node data of each node sample and the composite node data of its neighboring node samples.

2. The method according to claim 1, characterized in that, The graph neural network is used to obtain composite node data corresponding to the node sample based on the node data and behavior sample data of neighboring node samples associated with the same node sample, and to perform graph convolution operation on the composite node data of multiple node samples to obtain node update data of multiple node samples.

3. The method according to claim 2, characterized in that, The graph neural network is used to aggregate the composite node data of each node sample based on the normalized adjacency matrix to obtain the node aggregate data corresponding to the node sample. The neural network is used to fuse the node aggregate data of multiple node samples to obtain the node update data of multiple node samples. The normalized adjacency matrix is ​​determined by the relationship information of the node data.

4. The method according to claim 1, characterized in that, The graph neural network is a multi-layer graph neural network. The behavior preference prediction layer is used to determine the node target data corresponding to the node sample based on the node update data of the same node sample output by the multi-layer graph neural network. Based on the node target data of multiple node samples and multiple behavior sample data, the layer determines the multiple behavior preference prediction information of multiple user samples for different multimedia product samples.

5. The method according to claim 4, characterized in that, The target data for each node is determined by a weighted sum of the node update data corresponding to the node sample output by the multi-layer graph neural network.

6. The method according to any one of claims 1 to 5, characterized in that, The behavior prediction model is learned using a negative sampling-free learning method. Training the behavior prediction model based on the graph-structured sample data includes: The graph neural network obtains node update data for a node sample based on node data and behavior sample data of neighboring node samples associated with the same node sample. The behavior preference prediction layer determines multiple behavior preference prediction information for the corresponding user sample for the multimedia product sample based on multiple behavior sample data, multiple user sample node update data, and multiple multimedia product sample node update data. Extract multiple user samples with multiple behavioral sample preference labels for different multimedia product samples from the relational information of the node data; Based on the multiple user samples for different multimedia product samples, the multiple behavioral sample preference labels and multiple behavioral preference prediction information are used to determine the loss of the behavior prediction model. The behavior prediction model is updated based on the loss of the graph neural network.

7. A recommendation method, characterized in that, include: Input the target user data, target behavior data, and multiple multimedia product data into the behavior prediction model trained according to any one of claims 1 to 6; The behavior prediction model is used to obtain the target user's multiple behavior preference prediction information for different multimedia product samples from multiple user samples. If the target user's target behavioral preference information for the target multimedia product meets the recommendation criteria, the target multimedia product is recommended.

8. The method according to claim 7, characterized in that, The target user's target behavior preference information for the multimedia product includes the probability of the target user's target behavior for the multimedia product, and the recommendation conditions include: The predicted probability of the target user's target behavior towards the multimedia product is greater than or equal to a preset probability; and / or... The multimedia products are sorted in descending order of the predicted probability of the target user's target behavior towards the multimedia product. The target multimedia products include the first f multimedia products, where f represents an integer greater than or equal to 1 and less than the total number of multimedia products.

9. The method according to claim 7, characterized in that, The target behavior data includes multiple target behavior data, and the target user's target behavior preference index for the target multimedia product includes: the weighted result of the predicted probabilities of multiple target behaviors of the target user for the multimedia product.

10. A training device, characterized in that, include: The acquisition module is used to acquire graph structure sample data based on user basic information and heterogeneous user behavior information for multimedia products. The graph structure sample data includes node data of multiple node samples, multiple behavior sample data and relationship information of node data. The multiple node samples include user samples and multimedia product samples. The relationship information of node data includes heterogeneous behavior relationships between user samples and multimedia product samples. The training module is used to train a behavior prediction model based on the graph structure sample data. The behavior prediction model includes a graph neural network and a behavior preference prediction layer. The graph neural network is used to obtain node update data of the node sample based on the node data and behavior sample data of the neighboring node samples associated with the same node sample. The behavior preference prediction layer is used to determine multiple behavior preference prediction information of multiple user samples for different multimedia product samples based on multiple behavior sample data, multiple user sample node update data, and multiple multimedia product sample node update data. The graph neural network is used to obtain node update data corresponding to the user sample based on node data and behavior sample data of the multimedia product sample associated with the same user sample, and to determine node update data corresponding to the multimedia product sample based on node data and behavior sample data of the user sample associated with the same multimedia product sample. The data processing flow of the graph neural network is divided into a composite stage and a graph convolution stage. In the composite stage, the node data of each node sample is fused with the behavioral sample data associated with each node sample. In the graph convolution stage, the node data of each node sample is fused with the composite node data of each node sample and the composite node data of its neighboring node samples.

11. A recommendation device, characterized in that, include: The inference module is used to input target user data, target behavior data and multiple multimedia product data into the behavior prediction model trained according to any one of claims 1 to 6, and use the behavior prediction model to obtain the target user's multiple behavior preference prediction information for different multimedia product samples from the multiple behavior preference prediction information of multiple user samples for different multimedia product samples. The recommendation module is used to recommend the target multimedia product if the target user's target behavioral preference information for the target multimedia product meets the recommendation conditions.

12. An electronic device, characterized in that, include: processor; as well as, Memory for stored programs; The program includes instructions that, when executed by the processor, cause the processor to perform the method according to any one of claims 1 to 9.

13. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions for causing the computer to perform the method according to any one of claims 1 to 9.

Images