Social Recommendation Based On Trust And Matrix Factorization

With the rapid development of Internet and e-commerce,Internet information and resources are increasing at an unprecedented rate.The problem of information overload is becoming more and more serious.One of the hot issue in the era of big data is how to find information rapidly that users need.The recommender system can solve this problem of information overload effectively.Based on mining user historical behavior data and building accurate preference models for each user,it recommends information to users which may satisfy their needs.Nowadays,the serious problem faced by recommender system is data sparsity and cold-start.Using implicit feedback and social relationship can relief this problem.This thesis centered around the topic which is based on trust and matrix factorization.It explores how to fully exploit the trust relationship to help users obtain better personalized recommendation.In this paper,ranking recommender algorithm combining trust and similarity for implicit feedback and matrix factorization recommender algorithm incorporating user similarity and trust relationship are put forward separately.The main work of this thesis is as follows:Firstly,we summarize the research background,significance,the current research status at home and aboard,the definition and categories of traditional recommender system and advantages and disadvantages of these types of algorithms.We also introduce social recommender algorithms,two typical methods of Trust Measurement and three matrix factorization algorithms which provide theoretical basis for the research of this thesis.Secondly,existing algorithms only use explicit feedback data to recommend which faced with serious problem of data sparsity and cold-start.To solve this problem,this thesis proposes a ranking recommendation combining trust and similarity for implicit feedback from the perspective of ranking which research the impact of similarity and trust relationship for the result of ranking.First of all,this algorithms proposes mixture weight combining trust value and similarity to replace the original binary trust value.Then,it structure user's Characteristic matrix and realize trust propagation by weighting average trust neighbors' Characteristic matrices.At last,it optimizes users' preference for the items to get the ranking results of the candidate project directly.This paper also research the problem of measuring trust degree for the new user.It connect new users to trust network to relief the problem of cold-start of new users.Experiments show that thealgorithm has better performance compare to other baseline algorithms and it can solve the problem of data sparsity and cold-start effectively.Thirdly,Aiming at mostly trust-based recommender algorithms aren't distinguish the strength of trust relationship and only consider single similarity relationship or trust relationship,this paper propose a probabilistic matrix factorization algorithm incorporating user similarity and trust relationship according to user behavior characteristics.First of all,we create global trust degree by original trust network,and then measure local trust degree within the maximum propagation distance by trust propagation mechanism,then coordinate global degree and local degree to structure the final trust network and trust neighbors.At last,under the balance of user behavior characteristics,trust neighbors and similar neighbors modify users' characteristic matrices and integrate it into probabilistic matrix factorization model to shape efficient and credible model TSPMF.Experimental results on the public data sets show that this model can effectively improve the recommendation accuracy.