| With the development of Internet technology,the information in the network has also increased exponentially.While enjoying the rich content in the network,people also face the problem of information overload.Recommender system,as an information filtering technology,can assist people to find useful information from massive information.Collaborative filtering use user-item interaction information for personalized recommendation,but the recommendation effect is limited by data sparsity and cold start problems.With the development of online social platforms,more and more people participate in social activities and generate a large number of social relationships.Adding these social relationships to the recommender system,which can greatly alleviate the problems of data sparsity and cold start,brings new possibilities for the improvement of the recommendation system.However,the existing social recommendation algorithms are difficult to distinguish the user’s purchasing motivation,and cannot distinguish whether the user’s purchase is influenced by friends.Secondly,the social relationship information is too sparse,it is difficult to mine the high-order relationship of users from it,and it is impossible to make full use of the homogeneous relationship between users.Based on the above research questions,the main research of this paper is as follows:(1)We propose a Multi-channel Social Recommendation Model Based on Deep Clustering Grouping Strategy named MSDC.This model mainly includes three modules:deep clustering module,the aggregation module based on multi-channel graph attention network,and the rating prediction module.Among them,the deep clustering module is used to group users and items.The purpose of the deep clustering module is to divide the same type of items into a user-item interaction subgraph,and the same type of users into a item-user interaction subgraph,so as to complete the grouping of user interests.The aggregation module learns the attention of different sub-graphs related to prediction results,and its purpose is to mine users’ interests in different types of items and which type of users the item is more suitable for.The user embedding of each channel refers to the user’s interest in a certain type of item,so as to complete the decoupling representation of the user’s interest.The rating prediction module merges user embeddings and item embeddings into the multi-layer perceptron(MLP)to predict the ratings.Extensive experiments on multiple real-world datasets demonstrate that the proposed model MSDC is better than other social recommendation algorithms.Specifically,compared with the latest state-of-the-art model Graph Rec(Graph Neural Networks for Social Recommendation),the decrease of MAE(Mean Absolute Error)on the Ciao and Epinions datasets is 2.58% and 3.06%,and that of RMSE(Root Mean Square Error)is 2.26% and 2.07%,respectively.(2)We propose A High-Order Social Recommendation Model Based on Self-Supervised Learning named HSSL.This model firstly performs data augmentation to construct user collaboration graph and item collaboration graph,and performs self-supervised learning on these graphs to capture high-level user relationships and high-level item relationships.Secondly,the model constructs a neighbor-based supervision signal and performs neighbor-based self-supervised learning,so as to make full use of the homogeneity of users and the homogeneity of items,and reduce the noise in self-supervision.Extensive experimental results on three real-world datasets clearly demonstrate that the proposed model DSGM is better than other social recommendation algorithms.In particular,compared with the latest state-of-the-art model SEPT(Socially-Aware Self-Supervised Tri-Training for Recommendation),Precision@10 on Last FM,Douban-Book and Yelp datasets improves 1.97%,4.95%,and 4.93%,respectively,and Recall@10 improves 2.03%,5.55%,and 6.95%,respectively,and NDCG@10 improves 1.62%、5.46% and 7.47%,respectively. |
