Research On Recommendation Algorithm Based On User Social Relationship

A recommender system is a tool that can help users discover items of interest and effectively solve information overload.However,recommendation systems generally face the problems of data sparsity and cold start,especially when traditional recommendation algorithms such as collaborative filtering are used,users’ preferences for items are modeled only through user-item interaction records,which is easy to be disturbed by data sparsity and insufficient information,resulting in unsatisfactory prediction results.Because the interaction between users and items not only depends on their interests but also may be influenced by social friends.By recommending the items that the user’s friends like to the user,the data sparsity and cold start problems can be effectively alleviated,and the recommendation accuracy can be further improved.As a result,many studies on introducing social relationships into recommendation algorithms were born and gradually became a research hotspot.In addition,because the graph neural network has a powerful modeling ability for graph-structured data,the interaction relationships between users and items and the social relationships between users can be naturally expressed as graph-structured data.Therefore,this paper mainly uses the graph neural network as a modeling tool to conduct research on recommendation algorithms based on user social relationships.The main work and contributions of this paper are as follows:1.The session-based recommendation algorithm only captures the short-term dynamic interests of users,without considering the long-term interests of users and the influence of social friends on their behavior,resulting in poor accuracy and unable to achieve personalized recommendations.And most social recommendation methods assume that the influence of each friend is consistent when constructing a graph or modeling,which will easily introduce noise.Aiming at the above problems,a personalized recommendation algorithm combining social influence and long short-term preferences is proposed.Firstly,a novel heterogeneous relationship graph is designed to organize users’ social relationships and historical sessions,and a weighted pruning strategy is proposed to address the issue of inconsistent social influence that can easily introduce noise,reducing noise interference and enriching the graph structure information.Secondly,a heterogeneous graph neural network based on attention mechanism is proposed to learn graphs and obtain long-term preferences that integrate users’ social influence.Finally,the user’s short-term preference is captured using a lossless session modeling method,and the short-term preference and long-term preference are adaptively fused to obtain a feature representation that reflects the user’s global preference.The experimental results on Gowalla and Delicious datasets show that the accuracy of the proposed method on Hit Rate(HR)and Mean Reciprocal Rank(MRR)indicators is significantly improved compared with the existing advanced methods,which proves the effectiveness of the proposed algorithm.This algorithm fully captures users’ longterm and short-term preferences and social influence,and can accurately predict the next interaction item for users.Suitable for scenarios such as online shopping and Point-OfInterest(POI)Recommendations,and able to achieve recommendations with social attributes and personalization.2.Social recommendation algorithms based on graph neural networks often use different modules to model user-user graphs and user-item graphs separately,which not only ignores the association between items,but also hinders the transmission of information between nodes.Moreover,they also ignore the possible feature mismatch between the queried target user nodes and item nodes and their neighbor nodes,which leads to the introduction of noise and reduces the model performance.To solve the above problems,a social recommendation algorithm based on relational graph and dynamic neighborhood sampling is proposed.First,a relational graph containing multiple relationships between users and items is constructed,so that when the graph neural network is used to model the graph,it can capture the potential association between items and realize the cross-level transfer of information,and enhance the embedding representation of user nodes and item nodes.Then,a dynamic neighborhood sampling mechanism is designed to capture neighbor nodes with more consistent features with the target query pair,and reduce the noise interference of neighbors with inconsistent features when neighborhood aggregation.In the end,to further improve the prediction performance of the model,an enhanced graph neural network is proposed to model the relational subgraph obtained after sampling,and a more robust user and item embedding is obtained for scoring prediction.Tested on the Epinions and Ciao datasets,the prediction error of the proposed method on the Mean Absolute Error(MAE)and Root Mean Square Error(RMSE)indicators is significantly lower than that of the existing advanced methods,which proves the effectiveness of the proposed algorithm.This algorithm completes the rating prediction task in the recommendation system,aiming to accurately predict users’ ratings for items that have not been interacted with,in order to recommend content that they tend to score high,suitable for application to some content based online social and comment websites.3.The social recommendation algorithm based on the graph neural network often does not consider the interests of users and social friends,and the attractiveness of items will dynamic evolution.It can only capture the static features of users and items,which does not conform to the actual situation,and the recommendation performance is limited.For the problem,a social recommendation algorithm based on the global capture of dynamic,static,and relational features is proposed.First,an interaction modeling network and a temporal modeling network are designed to capture the long-term static representation and short-term dynamic representation of users and items,and a gated fusion network is used to adaptively fuse the two representations to obtain dynamic and static feature representations.Then,a relational aggregation network is designed to aggregate information from various relational neighbors to obtain relational feature representations.Finally,the dynamic and static feature representation and relational feature representations are fused to obtain a global feature representation,which realizes the global capture of dynamic,static,and relational features.By testing on two public datasets,the MAE and RMSE indicators are significantly improved compared with the existing advanced methods,which proves the effectiveness of the proposed algorithm.This algorithm combines the characteristics of the above two works and improves the shortcomings,fully capturing the dynamic,static,and relational features of users and items.In practical applications,it can not only predict user ratings for items more accurately,but also enhance the interpretability and application scenarios of the model.The research in this paper has successfully solved many deficiencies existing in the existing related algorithms.The proposed algorithm models not only have significant improvements in prediction accuracy and are suitable for various complex application scenarios,but also achieve recommendation effects with social attributes and personalization,which has broad application value.