Research On Recommendation Algorithm Using Social Network Analysis Combined With Collaborative Filtering Technology

With the rapid development of the Internet in recent years, information on the Internet became so much that it exceeds the effective limit of receiving, processing and using for an individual, which is known as the frequently mentioned information overload problem. Recommender system has been an important technology for dealing with the issue of information overload problem and many algorithms has been continuously proposed for it, many of which has achieved fairly good results in customized service, E-Commerce and computational advertising. However, despite of all the proposed algorithms, recommender system is still facing a number of challenges, among which are the difficulty of handling large-scale data, the inaccuracy of prediction, the sparsity of data, and the cold start problem. On the other side, with the advent of the social network services on the internet in the past few years, more and more users’ information in such network become usable. Some recommendation strategies has achieved remarkable results in using of such users’ information on certain social networks, but such information is not put into proper use in the current recommender system.This paper firstly introduced the current studies on the recommender system, analyzed and summarized the advantages and disadvantages of some existing collaborative filtering algorithms. Meanwhile, it introduced some methods in computing the node similarity in the social network analysis, analyzed and summarized their respective principles and application scenarios. On the basis of relevant outcomes and in view of the issues facing the current collaborative filtering technologies, it made an in-depth research on combining the social network analyses with collaborative filtering technologies and proposed some effective solutions.The work in this paper covers the following three aspects:(1) In view of information of users’social relationships on the social network, it firstly made the model of the social network, then it proposed the algorithm for calculating node similarity on social network in the method of machine learning. In this algorithm, it uses the information of the nodes themselves, the computed results of the traditional node similarity calculating algorithm and some auxiliary data as the training features, and uses the measures that whether there is an edge between node pairs as label, then uses the Logistic regression model as the training model to compute the node similarity. After computing the node similarity, it is proposed in this paper an algorithm with the node similarity data and the users’ rating data to predict the users’rating, which is referred to as the users’trustful rating in this paper.(2) In view of problems as excessive time and space occupations faced in the mass users recommender system in the collaborative filtering recommendation algorithm which leads to failure of making recommendation for new users, it is proposed in this paper an improved collaborative filtering recommendation algorithm which firstly builds a "users’characteristics-item" matrix, then solves the matrix with the latent factor model, and finally, on the basis of the solved matrix and users’characteristic data, predicts users’ratings on items by means of linear weightings, which is generally referred to as the users’preferential rating in this paper.(3) After computation of the users’trustful rating and preferential rating, the two ratings need to be inosculated so as to integrate a final users’rating on a certain item with a proper method. It is proposed in this paper a algorithm of inosculating the users’trustful rating and preferential rating in learning with machine, which uses the user’s trustful rating, preferential rating, number of neighbors on the social network and number of characteristics as the training features, and takes the user’s actual rating as the label, then uses the Fisher linear discrimination as the training model to predict user’s rating on certain items. The rating predicted with this algorithm is the user’s final rating on items.