A Recommendation Algorithm Based On Matrix Factorization That Integrates Trust Relationships And Item Popularity

With the rapid development of the Internet,the recommendation system has received more and more attention from people.Providing effective user personalized recommendation is a hot issue at present.Through the analysis of user historical behavior,user preferences are predicted.In recent years,the recommendation algorithm combined with user social network has been widely used.Relevant recommendation based on trust between users can effectively alleviate the problem of cold start in recommendation,but only based on the trust relationship between users will exist.Coverage issue.At the same time,due to the influence of network communication,media exposure and community discussion,items with higher popularity are more likely to be recognized by users,but the recommended coverage of unpopular items that are of interest to users is not high.In addition,there is an inherent trade-off between the coverage and accuracy of the recommended algorithm,which reduces the recommended accuracy while improving the recommended coverage.Aiming at the problem that the coverage of existing recommendation algorithms is not high,this paper proposes a matrix decomposition recommendation algorithm that combines item popularity and user trust relationship.Firstly,this paper defines the disconnection of trust network,and then proposes a recommendation algorithm TruMF based on the merged user-item scoring matrix and user-user trust relationship matrix.TruMF utilizes the transitivity of matrix decomposition technology,which regards user trust relationship and item score as the same level,which leads to the mixing of item score and trust relationship in the process of matrix decomposition,so that the transfer trust and prediction score occur simultaneously.This method greatly improves the coverage of the recommended algorithm,but loses the accuracy of the comparison method by about 8%.Aiming at this problem,this paper introduces the item popularity weighting strategy into the scoring matrix,and further optimizes it based on the TruMF algorithm,and proposes the PopTruMF algorithm.The PopTruMF algorithm studies the problem of learning the matrix decomposition model from implicit feedback.Compared with the previous work of giving uniform weight to the lost data,this paper personalizes the lost data according to the popularity of the item,and weights the user's scoring item and unscoring item separately.At the same time,an incremental update strategy is designed to adjust dynamic data in real time to meet the requirements of online learning and effectively learn from implicit feedback.Through a series of offline and online protocol comparison experiments,it is proved that the PopTruMF algorithm can greatly improve the recommended coverage rate while ensuring the accuracy of recommendation,and can give users better recommendation results.