Research On The Personalized Recommendation Based On Tag Clustering And Trust Relationship In Social Network

As one of the most important application in the of Web2.0,social network like Facebook,Weibo,etc.is such a place where people spend tremendous amount of time.Many of them have over billions of day active user leading to the information overload,which causes that users hardly find the information they want instantly.On the other hand,for these information producers,their content is much harder than before to stand out and get prevailing.If the social network provider cannot address these problems on effective information matching,the user satisfaction and hit rate are bound to decline.To solve the information asymmetry,personalized recommendation system emerged.It is allowed to connect users with resources to help them identify valuable data.Due to the exaggerated amount of information in social network,it's difficult for traditional recommendation model to integrate users and resource,which thus impact the accuracy of recommendation and furthermore the user satisfaction.But UGC tag and trust,the key sources in social recommendation are good at reflecting user similarity on interests and resource features on preference,favor of information organization and accuracy rise.However,sparse problem stands out because of the high-dimension on UGC tag.Generally,the tensor decomposition model suffices to cope with sparsity,but it will bring semantic loss when calculating the tag similarity on 3-dimension,resulting to accuracy reduction.On the other hand,in terms of trust,single faceted trust fails to indicate the real trust strength between users because on the social network people's trust is often built on one or more fields of interests.Therefore,it's essential to construct multi-faceted trust to identify the trust strength on different fields,which can precisely get the user similarity and enhance recommendation accuracy.The problem is that in some social network,it's hard to identify different facets due to the numerous resources.According the above,we first introduce a spectral clustering method based on co-occurrence individual and group similarity to cluster UGC tag.Then it is incorporated with tensor factorization by improving the initial tensor to handle the problem of semantic loss.Not only can this method reduce sparsity to increase the recommendation accuracy,but also it can help diffuse the user preference who will receive a variety of resources in order to better the recommendation diversity.In the experiment part,we simulated the proposed two models on the real dataset from Last.fm.Briefly,the first one is called Co-occurrence Spectral Clustering with inverse document frequency(Co SClu IDF)and the second is called Multi-Faceted Trust Co SClu IDF(MTF-Co SClu IDF).Each of them is compared with several other methods on recommendation accuracy and diversity.The results show that Co SClu IDF is able to completely hold the whole ternary semantic relation in tackling tag redundancy and sparsity.Thus,the recommendation accuracy is significantly enhanced with a certain increase on diversity.Meanwhile,MFT-Co SClu IDF can further lift the recommendation accuracy without diversity decline,which means the method limitedly improves diversity in line with the former research.