Research And Implementation Of Social Recommendation Algorithm Based On Trust Relationship

With the rapid development of information society,the data on the Internet has been growing explosively,showing the characteristics of a wide variety and complex structure,and traditional recommendation algorithms cannot effectively solve the problem of information overload.With the emergence of social applications,recommendation technology that integrates social information has begun to become a research focus in academia and industry.Social recommendation algorithms integrate the trust information of users' social network,which can effectively alleviate problems of data sparsity and cold start.However,most social recommendation algorithms only model item preference of users themselves,ignoring the preference information of users' trusted friends,and do not fully explore the transmission of user information on social networks.The key of social recommendation algorithm is how to utilize trust relationship effectively to mine deep-level user and friends' preference information and interaction relationships.Therefore,this paper focuses on the research of social recommendation algorithm based on trust relationship,mainly doing the following work:(1)To solve the problem that traditional social recommendation algorithms do not use trust relationship effectively to mine user friend preference information,a graph memory network model(PA-GMN)integrating trust relationship and user preference is proposed firstly.This paper first uses attention mechanism to integrate user preference information of different rating levels in user latent vector learning,and secondly uses memory network to store multi-aspect shared preferences of users and their friends and uses graph neural networks to model user trust relationships and preferences in social networks information transferability,then finally uses user latent vector and item latent vector for rating prediction.Experiment compared with other social recommendation algorithms,shows that this model can effectively improve the prediction accuracy on the existing open source dataset.The results show that the PA-GMN algorithm further alleviated problems of data sparsity and cold start,and improve the performance of recommendation systems.(2)As PA-GMN algorithm cannot distinguish the influence of multi-aspect preferences of users and friends to an item,this paper proposes an improved self-attention mechanism-based multi-aspect preference graph attention network model PA-GAN.This model mainly uses self-attention mechanism to capture different influences of friends' various preferences on items,then weights and aggregates user's trust relationship from different friends to generate friend vector that incorporate preference information to train the final user latent vector,and finally uses user latent vector and item latent vector for rating prediction.Experimental results show that the proposed model PA-GAN can greatly improve the accuracy of recommendation.Compared with PA-GMN algorithm,this model also has a great improvement in recommendation performance.(3)Based on PA-GMN and PA-GAN algorithms that integrate trust relationships and user multi-aspect preferences,we design and implement a social recommendation system.The system includes data preprocessing module,graph topology building module,model training module,data storage module,and user recommendation module.This platform is mainly responsible for visualization of data processing,modeling training and showing recommendation results.