Recommendation method and device based on social network structure information

Abstract

The present invention discloses a recommendation method and device based on social network structural information. The method includes: constructing a user project view, a co-purchase view, and a social friend view based on decoupling relationships; calculating the structural similarity of social groups in the co-purchase view and the social friend view; presetting semantic vectors in the co-purchase view and the social friend view, respectively, and calculating the semantic similarity of users in the co-purchase view and the social friend view; generating real data based on structural similarity and pseudo data based on semantic similarity, distinguishing real data from pseudo data so that semantic similarity approaches structural similarity; and updating the user representation in the user project view based on semantic information from the social friend view and the co-purchase view to predict user interests. The present invention has better performance in social recommendation tasks and improves the accuracy of recommendations.

Description

Technical Field

[0001] The present invention relates to the field of information recommendation, and in particular to a recommendation method and device based on social network structure information. Background Art

[0002] The rapid development of internet technology has simultaneously presented a vast amount of information to the public. Traditional search algorithms can only present the same ranking results to all users, failing to tailor services to their individual interests. This information explosion has actually reduced information utilization. Furthermore, with the proliferation of social media, users have developed rich social connections through platforms like WeChat, Weibo, Douban, Zhihu, and TikTok. More and more people are willing to share their opinions on various items on social media, forming a vast social network among users. Furthermore, users' attitudes and evaluations of certain things are significantly influenced by their friends, and they often share similar interests and hobbies with their friends. Therefore, it is crucial to leverage information from social networks to improve the accuracy of recommendation systems and alleviate the problem of information overload.

[0003] One type of recommendation model in the existing technology is based on social networks, which is mainly based on Gaussian matrix decomposition or Poisson matrix decomposition, and integrates social network information into traditional matrix decomposition methods. Among them, some people propose to use trust metrics to replace collaborative filtering methods to recommend items that users may like. Trust metrics refer to algorithms that can propagate trust on trust networks and estimate trust weights that can be used to replace similarity weights; some people assume that similar users may share similar implicit preferences, and thus incorporate social network information into the model by regularizing the target user's latent variables with the latent variables of friends; some people propose the TrustSVD method, which uses the preferences of friends as auxiliary implicit feedback for the target user, thereby integrating the influence of the target user's friends into SVD++; some people propose the SocialMF model, which uses the trust relationship between users as social network information and takes the influence of the user's friends into account in the matrix decomposition model; some people assume that users tend to assign items that their friends like. Higher ratings are obtained and this assumption is incorporated into the pair-wise ranking loss function, and the SBPR model is proposed; a new method is proposed to integrate dual sparse information, namely the regular rating data given by users and the social trust network between users to improve the performance of collaborative filtering recommendations; a joint probability generation model is proposed to integrate multiple factors such as geographical factors, time effects, social relevance, content information and popularity influence to deal with data sparsity problems; the SERec method is proposed to simulate users' exposure to items, and then use these exposures to guide rating predictions, integrating social exposure into collaborative filtering, which is less restricted and closer to reality; a general framework is proposed to integrate three commonly used CF methods, and incorporate social information into the model through graph embedding, which is finally expressed as SoTriCF.

[0004] Another approach is to apply deep learning to graph neural networks for graph data. Some have used graph convolutional neural networks to capture the process by which user preferences are influenced by their social networks, and proposed the socialGCN model; others have proposed the GraphRec model, which uses graph neural networks to aggregate implicit representations of users and items from their neighborhoods, while incorporating vector representations of users' opinions on items to solve the task of rating prediction; others have propagated vector representations on a bipartite graph composed of users and items to encode high-order connectivity on the bipartite graph; others have used hierarchical propagation structures to model recursive dynamic social diffusion in social recommendations; and others have proposed a model that focuses on modeling the indirect influence of high-order neighbors in social networks. , to improve the performance of item recommendation, social relations are directly considered in the prediction model, aiming to learn better user embeddings to improve recommendations; it is proposed to integrate explicit and implicit influences in the GCN network to learn the social embeddings of users and items in social recommendations, and the experimental results preliminarily prove the effectiveness of the model; a multi-channel hypergraph convolutional network is proposed, which is suitable for multi-modal hypergraphs, and self-supervised learning is integrated into training to improve information loss during feature fusion; the SEPT training framework is proposed, which is a socially aware self-supervised triple training framework that improves recommendation accuracy by generating self-supervised signals from two complementary views of the original data.

[0005] However, both of the above methods have disadvantages:

[0006] Most matrix factorization-based methods incorporate social network information into the model in the form of regularization terms in the loss function, which has certain limitations. Furthermore, when predicting the final recommendation results, shallow functions such as cosine similarity are used, making the model's ability to predict user preferences relatively insufficient.

[0007] Most social recommendation algorithms based on graph data only consider the semantic information of user nodes in the graph. However, the local structural information of nodes in the graph also has a very important influence on the feature representation of user nodes. Therefore, the feature representation that only contains semantic information has a great impact on the final recommendation results. Summary of the Invention

[0008] The present invention aims to solve one of the technical problems in the related art at least to a certain extent.

[0009] Therefore, the purpose of the present invention is to solve the problem of how to use social network information to improve the accuracy of the recommendation system to alleviate the problem of user information overload, and proposes a recommendation method based on social network structure information.

[0010] Another object of the present invention is to provide a recommendation device based on social network structure information.

[0011] To achieve the above objectives, the present invention proposes a recommendation method based on social network structure information, comprising the following steps:

[0012] Based on the decoupled relationships, construct user project views, common purchase views, and social friend views;

[0013] calculating structural similarities of social groups in the co-purchase view and the social friend view;

[0014] Presetting semantic vectors in the co-purchase view and the social friend view respectively, and calculating semantic similarity of the user in the co-purchase view and the social friend view;

[0015] generating real data according to the structural similarity, generating mimetic data according to the semantic similarity, distinguishing the real data from the mimetic data so that the semantic similarity approaches the structural similarity; and

[0016] The representation of the user in the user-item view is updated by combining semantic information of the social friend view and the co-purchase view to predict user interests.

[0017] According to an embodiment of the present invention, a recommendation method based on social network structural information can learn the structural information of users from different perspectives and integrate the sub-network structure of social groups and their semantic collaboration information into a unified recommendation framework. By learning social influence from the network structure, it achieves better performance in social recommendation tasks, fully learns the features of users in different views, integrates useful information, and ultimately improves the accuracy of recommendations.

[0018] In addition, the recommendation method based on social network structure information according to the above embodiment of the present invention may also have the following additional technical features:

[0019] Furthermore, the user-item view is composed of interacting user and item nodes and edges of the interactive relationships between users and item nodes; the edges of the co-purchase view are established between two users; and the edges of the social friend view are constructed based on social friend relationships.

[0020] Furthermore, the calculating of the structural similarity of social groups in the co-purchase view and the social friend view includes:

[0021] The co-purchase view and the social friend view are preset, and the normalized probability of the neighbors of the user is calculated according to the adjacency matrix in the co-purchase view.

[0022] Furthermore, the method further includes: based on the semantic similarity between any two views, merging semantic information of other views to update the representation of the user in the co-purchase view.

[0023] Furthermore, when the structural similarity is calculated based on the substructure of the social group, the output logarithmic probability is maximized, and when the structural similarity is calculated based on the substructure of the social group, the output logarithmic probability is minimized.

[0024] Furthermore, the updating of the user representation in the user-item view in combination with the semantic information of the social friend view and the co-purchase view to predict user interests includes:

[0025] A first user and a first item are preset, and the first user and the first item are optimized and trained using Bayesian personalized ranking to calculate the rating of the second user for the first item.

[0026] To achieve the above-mentioned object, the present invention further proposes a recommendation device based on social network structure information, comprising:

[0027] A construction module for constructing user-item views, co-purchase views, and social friend views based on decoupled relationships;

[0028] A first calculation module is used to calculate the structural similarity of social groups in the common purchase view and the social friend view;

[0029] a second calculation module, configured to preset semantic vectors in the co-purchase view and the social friend view, respectively, and calculate semantic similarity of the user in the co-purchase view and the social friend view;

[0030] a generating module, configured to generate real data according to the structural similarity, generate mimetic data according to the semantic similarity, and distinguish the real data from the mimetic data so that the semantic similarity approaches the structural similarity; and

[0031] A prediction module is configured to update a representation of a user in the user-item view by combining semantic information of the social friend view and the co-purchase view to predict user interests.

[0032] According to an embodiment of the present invention, a recommendation device based on social network structural information can learn the structural information of users from different perspectives and integrate the sub-network structure of social groups and their semantic collaboration information into a unified recommendation framework. By learning social influence from the network structure, it achieves better performance in social recommendation tasks, fully learns the features of users in different views, integrates useful information, and ultimately improves the accuracy of recommendations.

[0033] Beneficial effects of the present invention:

[0034] The present invention can learn the structural information of users from different perspectives and integrate the sub-network structure of social groups and their semantic collaboration information into a unified recommendation framework. By learning social influence from the network structure, it achieves better performance in social recommendation tasks, fully learns the features of users in different views, integrates useful information, and ultimately improves the accuracy of recommendations.

[0035] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0037] Figure 1 is a flowchart of a recommendation method based on social network structure information according to an embodiment of the present invention;

[0038] Figure 2 A schematic diagram of a framework of a recommendation method based on social network structure information according to an embodiment of the present invention;

[0039] Figure 3 2 is a schematic diagram of the structure of a recommendation device based on social network structure information according to an embodiment of the present invention. DETAILED DESCRIPTION

[0040] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0041] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0042] The following describes a recommendation method and apparatus based on social network structure information according to an embodiment of the present invention with reference to the accompanying drawings. First, the recommendation method based on social network structure information according to an embodiment of the present invention will be described with reference to the accompanying drawings.

[0043] Figure 1 The figure is a flowchart of a recommendation method based on social network structure information according to an embodiment of the present invention.

[0044] like Figure 1As shown, the recommendation method based on social network structure information includes the following steps:

[0045] Step S1: construct a user project view, a common purchase view, and a social friend view based on the decoupling relationship.

[0046] Specifically, to avoid negative information transfer across heterogeneous views, different relationships are decoupled into three views: the "user-item" view (user-item relationship), the "co-purchase" view (user-user relationship), and the "social-friend" view (user-user relationship). Different views are aligned by users who appear in the view at the same time. Useful social information is first selected from the homogeneous "co-purchase" and "social friend" views, and then integrated into the "user-item" view to ultimately predict user interests. The details are as follows:

[0047] User-item view: consists of interacting user and item nodes and the edges of their interaction relationships;

[0048] Co-purchase view: An edge in the co-purchase view is created between two users if they purchased the same item. This intuitively reflects the common purchasing habits between users. Considering that the strength of co-purchase relationships is low and may introduce noise to social influence learning, we filter the edges between two users by adding the constraint of social friend relationships.

[0049] Social Friend View: The edges in the social friend view are constructed based on social friend relationships. Considering that some friends may occasionally establish relationships, this does not contribute significantly to modeling a user's social information. Based on relationship stability: The friend structure in the ternary closure is stable, and the edges of the ternary closure social structure are preserved in the social friend view.

[0050] Step S2: Calculate the structural similarity of social groups in the co-purchase view and the social friend view.

[0051] Specifically, we generate real training instances by computing the structural similarity of social groups in the co-purchase and social friend views. Given the co-purchase view C and social friend view S, for a user, we compute the normalized probability of its neighbors based on the adjacency matrix MC in the co-purchase view.

[0052] Step S3: Preset semantic vectors in the co-purchase view and the social friend view respectively, and calculate the semantic similarity of the user in the co-purchase view and the social friend view.

[0053] Specifically, imitation data is generated based on the similarity between the user's semantic representations in two views. Given a specific view C, a graph neural network is used to capture the user's semantic information by aggregating neighborhood information in the view, which is called "intra-view aggregation".

[0054] The semantic representation of a user in the social friend view can be calculated in the same way. Given the semantic vectors in the co-purchase view and the social friend view, the semantic similarity of the user in the two views is calculated.

[0055] Based on the semantic similarity of the two views, the representation of users in the co-purchase view C (the same users in the social friend view S) is updated by incorporating information from other views, which is called “cross-view aggregation”.

[0056] Step S4: generating real data based on the structural similarity, generating mimetic data based on the semantic similarity, distinguishing the real data from the mimetic data, and making the semantic similarity close to the structural similarity.

[0057] Specifically, real data and mimic data are generated based on structural and semantic similarity, respectively. By distinguishing between real data and mimic data, the semantic similarity of user representations is forced to approach the inherent structural similarity, thereby injecting structural information into the user's semantic representation learning.

[0058] Following the optimization of the GAN, the output log probability is maximized when the similarity is calculated by the substructure of the social groups, and minimized when the similarity is calculated based on the substructure of the social groups.

[0059] Step S5: updating the user's representation in the user-item view by combining the semantic information of the social friend view and the co-purchase view to predict the user's interests.

[0060] Specifically, the user representation in the user-item view is updated by combining semantic information from the social friend view and the co-purchase view to make the final prediction. Given a user u and an item i, Bayesian personalized ranking (BPR) is used for optimization training to calculate the rating of user u for item i.

[0061] Figure 2 3 is a schematic diagram of a framework of a recommendation method based on social network structure information according to an embodiment of the present invention.

[0062] Through the above steps, the present invention can learn the structural information of users from different perspectives, and integrate the sub-network structure of social groups and their semantic collaboration information into a unified recommendation framework. By learning social influence from the network structure, it achieves better performance in social recommendation tasks, fully learns the features of users in different views, integrates useful information, and ultimately improves the accuracy of recommendations.

[0063] It should be noted that there are many ways to implement recommendation methods based on social network structure information. However, no matter what the specific implementation method is, as long as the method can use social network information to improve the accuracy of the recommendation system, it is a solution to the existing technical problems and has corresponding effects.

[0064] In order to implement the above embodiment, Figure 3 As shown, this embodiment also provides a recommendation device 10 based on social network structure information. The device 10 includes: a construction module 100, a first calculation module 200, a second calculation module 300, a generation module 400, and a prediction module 500.

[0065] A construction module 100 is used to construct a user project view, a common purchase view, and a social friend view based on the decoupled relationship;

[0066] A first calculation module 200 is used to calculate the structural similarity of social groups in the co-purchase view and the social friend view;

[0067] A second calculation module 300 is configured to preset semantic vectors in the co-purchase view and the social friend view, respectively, and calculate semantic similarity of the user in the co-purchase view and the social friend view;

[0068] A generating module 400 is configured to generate real data based on structural similarity, generate mimetic data based on semantic similarity, and distinguish between real data and mimetic data so that the semantic similarity approaches the structural similarity; and

[0069] The prediction module 500 is used to update the representation of the user in the user-item view by combining the semantic information of the social friend view and the co-purchase view to predict the user's interests.

[0070] It should be understood that in the construction module 100, the user-item view is composed of interacting user and item nodes and edges of the interactive relationships between users and item nodes; the edges of the co-purchase view are established between two users; and the edges of the social friend view are constructed based on social friend relationships.

[0071] It should be noted that the first calculation module 200 is also used to preset the co-purchase view and the social friend view, and calculate the normalized probability of the user's neighbors based on the adjacency matrix in the co-purchase view; the prediction module 500 is also used to preset the first user and the first item, and use Bayesian personalized ranking to optimize the training of the first user and the first item to calculate the second user's score for the first item.

[0072] According to an embodiment of the present invention, the recommendation device based on social network structural information can learn the structural information of users from different perspectives, and integrate the sub-network structure of social groups and their semantic collaboration information into a unified recommendation framework. By learning social influence from the network structure, it achieves better performance in social recommendation tasks, fully learns the features of users in different views, integrates useful information, and ultimately improves the accuracy of recommendations.

[0073] It should be noted that the above explanation of the embodiment of the recommendation method based on social network structure information is also applicable to the recommendation device based on social network structure information of this embodiment, and will not be repeated here.

[0074] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0075] In the description of this specification, the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0076] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A recommendation method based on social network structure information, characterized in that: The following steps are involved: Based on the decoupled relationships, a user-item view, a co-purchase view, and a social friend view are constructed. In the co-purchase view, edges are established between two users if they purchased the same item, reflecting the common purchasing habits of the users. In the social friend view, edges are constructed based on social friend relationships. Calculating structural similarities of social groups in the co-purchase view and the social friend view; Presetting semantic vectors in the co-purchase view and the social friend view respectively, and calculating semantic similarity of the user in the co-purchase view and the social friend view; generating real data according to the structural similarity, generating mimetic data according to the semantic similarity, distinguishing the real data from the mimetic data so that the semantic similarity approaches the structural similarity; and The representation of the user in the user-item view is updated by combining semantic information of the social friend view and the co-purchase view to predict user interests.

2. The method according to claim 1, characterized in that The user-item view consists of interacting user and item nodes and edges of the interactive relationships between user and item nodes; the edges of the co-purchase view are established between two users; The edge of the social friend view is constructed according to the social friend relationship.

3. The method according to claim 1, characterized in that The calculating the structural similarity of social groups in the co-purchase view and the social friend view includes: The co-purchase view and the social friend view are preset, and the normalized probability of the neighbors of the user is calculated according to the adjacency matrix in the co-purchase view.

4. The method according to claim 1, wherein The method further includes: based on the semantic similarity between any two views, merging semantic information of other views to update the representation of the user in the co-purchase view.

5. The method according to claim 1, wherein When the structural similarity is calculated from the substructure of the social group, the log probability is maximized, and when the structural similarity is calculated based on the substructure of the social group, the log probability is minimized.

6. The method according to claim 1, characterized in that The updating of the user representation in the user-item view by combining the semantic information of the social friend view and the co-purchase view to predict user interests includes: A first user and a first item are preset, and the first user and the first item are optimized and trained using Bayesian personalized ranking to calculate a rating of the first user for the first item.

7. A recommendation device based on social network structure information, characterized in that: include: A construction module is configured to construct a user-item view, a co-purchase view, and a social friend view based on the decoupled relationships. The co-purchase view includes edges between two users, where the edges reflect the same purchasing habits of the users if they purchased the same item. The social friend view includes edges based on social friend relationships. A first calculation module is used to calculate the structural similarity of social groups in the common purchase view and the social friend view; a second calculation module, configured to preset semantic vectors in the co-purchase view and the social friend view, respectively, and calculate semantic similarity of the user in the co-purchase view and the social friend view; a generating module, configured to generate real data according to the structural similarity, generate mimetic data according to the semantic similarity, and distinguish the real data from the mimetic data so that the semantic similarity approaches the structural similarity; and A prediction module is configured to update a representation of a user in the user-item view by combining semantic information of the social friend view and the co-purchase view to predict user interests.

8. The device according to claim 7, characterized in that The user-item view consists of interacting user and item nodes and edges of the interactive relationships between user and item nodes; the edges of the co-purchase view are established between two users; The edge of the social friend view is constructed according to the social friend relationship.

9. The device according to claim 7, characterized in that The first calculation module is further configured to: The co-purchase view and the social friend view are preset, and the normalized probability of the neighbors of the user is calculated according to the adjacency matrix in the co-purchase view.

10. The device according to claim 7, characterized in that The prediction module is further used to: A first user and a first item are preset, and the first user and the first item are optimized and trained using Bayesian personalized ranking to calculate a rating of the first user for the first item.

Images