The Research On Personalized Recommender System Based On Matrix Factorization

With the rapid development of social networks,many users spend more and more time on social networking applications.As we all know,a large number of userful data exists on social platforms which includes microblogging short text,comments between users and user registration information.If we go to research in this part of the data we will find a lot of content has been published emotionally,such as evaluating an item with a positive or negative aspect and expressing people's demand for some goods.Actually hundreds of millions of users are active on the microblogging,the text data hide a lot of valuable information they produced.In this paper,that research on personalized recommender system based on microblogging is a very valuable research topic.The traditional recommender system can not solve the complex problem of social networks recommended.The main difficulties are the following,first the context information has a great influence on the performance of the recommender system on social platforms,second user's interest preferences changed occur over time in social networks,and then the user interest model requires a large number of short text content,finally need to achieve efficient and accurate recommendation algorithm etc.In this paper,we proposed a hybrid prediction model that based on matrix factorization algorithm,and then we changed the recommend problem into a matrix factorization problem with this way.Finally according to the prediction matrix to achieve articles recommended.The main innovation of this paper include three parts,first of all,we proposed a recommender system architecture that can be divided into two modules which the one is a user modeling module and the other is a recommendation engine module.And then in this paper on user modeling module the work includes microblogging crawling system,text processing system,an incremental update user model system based on Rocchio algorithm,these work realized from the acquisition,processing,content extraction,incremental updating of data and other steps to the greatest extent reflect the user's preference for dynamic interest.Finally in the recommendation engine module we proposed a hybrid prediction model fusion context data based on matrix factorization model,these context data includes demographic information,a collection of mutual concern to friends,the consumer records of users.In this paper,we model different types of data in the same hybrid prediction model by different methods.Full use of various types of contextual information optimizes the performance prediction model.Obtaining the users and items porential feature prediction matrix according to learn parameters of the hybrid prediction model.So we can achieve the high accuracy items recommended with this model.After experiments we found the result of serveral evaluation the hybrid prediction model is superior to collaborative filtering algorithm and general SVD algorithm.In this paper of incremental changes to get the user model to achieve a dynamic user interest preference,and in the matrix factorization model added the positive and negative feedback,the sciring matrix split into two matrices which called positive feedback matrix and negative feedback matrix,we could bidirectional extrack the potential user preferences.The experimental verification of the matrix model based on positive and negative feedbacks can greatly improve the quality of recommendation.In this paper the study shows that the social recommended system in the context of social network and the users' positive and negative feedback data are the most important factor to improve the recommended quality.