Research On Social Based Generative Model And Recommender System

Recommender system is able to mine the resources that may interest users and beneficial to alleviate the information overload problem.Incorporating social network information into recommender system not only meets users' selecting patterns in real world,but also alleviates the sparseness and cold start problem in recommender system.Therefore,social based recommender system has caused in-depth discussion not only in academia but also in industry.There are several problems in most existing social based recommender system.Firstly,the core of the social based recommender system lies in calculating the influence between users.Most algorithms obtain the similarity between users by using similarity function,and regard it as the social influence.It is no doubt that different similarity function results in different recommendation performance.Secondly,most algorithms do not simulate user's product selection process in real world,which neglect the dynamic procedure that user's preferences and social networks interact with each other.Based on the above problems,this paper proposes two social based recommendation models,SoGeM(social based generative model for recommendation)and SoMF(social based matrix factorization model for recommendation).We conduct massive experiments and the results show that these two models work better than other 6 classic baseline recommendation algorithms.The main achievements of this paper are as follows:(1)We propose a social based recommendation model SoGeM,which is based on LDA.SoGeM assumes that user's interaction with items comes from user's own preferences or influence from social friends,and user's purchase behaviors is simulated as a generative process.We use Gibbs sampling method to train SoGeM,and user's purchase pattern and social influence between friends can be captured through the learning process.The probability that user will buy never-purchased item can be calculated through the learned parameters and recommend users the top-n items.(2)This paper describes the intrinsic relationship between LDA and MF.Moreover,the hypothesis of SoGeM is applied into the MF and SoMF model is proposed.We use SGD(stochastic gradient descent)method to train SoMF model,and the user's purchase pattern and social influence between friends also can be captured automatically,which is used to make recommendations to users.(3)We conduct experiments on 4 datasets extracted from different social platforms to do performance verification,and the effectiveness is evaluated by using 3 well-known metrics,precision,recall and MRR(mean reciprocal rank).Then we design and implement a prototype system based on the proposed 2 models.This paper first shows the significance of research on social based recommender systems and points out the problems in most existing recommendation approaches.Then we propose two social based recommendation models,SoGeM and SoMF.We conduct experiments to validate the effectiveness of the proposed two models and design the prototype system.Finally we summarize our work and plan the future researches.