Research On Social-based Recommender System

As the famous saying goes “Birds of a feather flock together”,users' beahaviors will be affected by their trusted friends.It's easier to accept friends' opinions than the advertisement.Combining traditional recommendation algorithms with social network can not only improve the accuracy of result but also alleviate the cold start and sparsity problems in recommendation systems.However,social-based recommender systems still have some problems.First,these algorithms unilaterally use user similarity to illustrate user influence and ignore the importance of user influence.Second,these algorithms are generally extended on collaborative filtering and matrix factorization.Collaborative filtering methods suffer from sparsity,cold start problems.When the input data become complex,it's hard for matrix factorization to consider all relationship between the attributes.In order to solve the above problems,we propose a social recommender system with factorization machines called SocialFM to improve the quality of recommendation and user satisfaction.The main researches and contributions of this thesis are as follows:1.This paper researches and analyzes the characteristics of social networks,sums them up as three characteristics: different,dynamic and weak transitive.Then we propose two methods to compute the users' social influence.The first method calculates the degree of user expertise,social connection and similarity between users.Then combine these three measures to calculate the influence of the users.The second one is to use the SGD method to learn the influence.2.This article propose a social-based factorization machines called SocialFM.This model utilizes the item tag or LDA algorithm to calculate the item feature vector.Moreover,use this vecotr and user rating records to calculate user feature vector and user similarity.Then,based on the factorization machines,the input data is transformed into the vectors with the user behavior as the principal,the friend relationship,the historical score record and the item feature as auxiliary.Afterwards,we propose social network regularization and the model internal regularization to constrain the objective function.Finally,we use stochastic gradient descent algorithm to train the model and obtain the final recommendation result.3.We use the Douban data set to conduct experiments.We use MAE and RMSE to evaluate result.The experiment results show that our approach outperforms other state-of-the-art recommendation methods.Then we compare and analyse four different methods about user influence.At last,based on above research,we design and implement a social-based sharing and review system.In this paper,we first describe the background and significance of social-based recommender system.In the second charpter,after analyzing some social-based recommender algorithms and their shortcomings,we put forward the framework of technical route.In the third charpter,we introduce the key algorithms in SocialFM in detail.Afterwards,experimental results are illustrated int fourth charpter.The experimental results show that our approach outperforms other state-of-the-art methods.In the fifth charpter,we implement a prototype system based on these ideas.At last,we summarize this paper and discuss the future work.