Research On Recommendation Methods Based On User Roles And Tensor Decomposition

With the popularity of the Internet,the Internet service platform based on interactive behavior has achieved unprecedented development.Social network has become an important communication platform in people's daily life through a large number of information sharing and views expression.The growing information resources caused the problem of information overload,and the emergence of the recommendation system to alleviate the impact of information overload.Collaborative filtering recommendation technology is currently the most widely used recommendation method.However,there are still inadequacies in collaborative filtering algorithms,the most obvious of which is data sparsity and difference insufficient of recommendation.This thesis summarizes the research status of recommendation systems in social networks in recent years.According to the analysis of user behaviors,user role discovery and tensor decomposition improvements are studied.The main research work is as follows:1.This thesis designs a role discovery model based on entropy principle.Considering the massive user behavior data in social networks,the user roles are measured from three dimensions based on the information entropy principle.Firstly,based on the user's behavioral preference in social network,the user's interest diversity is caputured by using information entropy.Secondly,considering the user's social relations and ability of information dissemination,the entropy weight method is used to comprehensively calculate the social influence of users.Finally,analyze the user's participation behavior in different periods and use the cloud model to transform the characteristics of the sequence data,then the entropy of the cloud model is calculated the user's activity level.To sum up,it's performed the role discovery for users from different perspectives in reasearch,and combined with the collaborative filtering algorithm to enhance the recommendation effect.2.This thesis proposes a recommendation model based on dynamic role identification and tensor decomposition.Firstly,information entropy is used to capture user roles for solving the problem of identifying indiscriminate user roles.Secondly,considering the drifting phenomenon of user interests,time window is proposed to describe the dynamic role identification,which solves the problem of non-preference difference of individual user generated by static role identification.Finally,a rating prediction model based on "user-item-role" tensor decomposition is constructed,which could improve the performance of rating prediction.And the characteristics of tensor in data transformation is introduced into the model.In addition,during the tensor decomposition,accuracy of rating prediction is improved by iteratively dealing with the missing value,and recommendation list for target user is generated.Finally,this thesis uses the real data of douban to verify and analyze the model.Experiments show that the method to alleviate the problem of rating prediction inaccurate without role identification,and can effectively improve performance of recommendation in the environment of sparse data.