Social Recommendation Algorithm Research Based On User Trust Influence

With the rapid development of the mobile Internet, we media and social networks, information overload has been emerged in the Internet. As a revolutionary technology after the search engine, the recommendation system could discover the potential needs of users and improve the efficiency of people’s selection and filtering items, which make it become the focus of industry and research fields. As an important algorithm of recommendation system, Collaborative filtering Algorithm merely depends on the user rating information, whose principle is simple but efficient. However, it is in trouble when meets cold start and data sparsity.Matrix decomposition and the introduction of social network information are methods to solve the problem of cold start and data sparsity. Matrix decomposition means transform the original score matrix information to the product of two low rank matrix, then utilize the low rank sub matrix to fit the original rating matrix, which will change score prediction problem into the optimization problem with an object function. On the one hand, we can solve the users and items latent factors through the gradient descent, on the other hand introduced more ancillary information that is associated with the users and items to regularize users and items feature vector, resulting in more accurate prediction models. Social recommendation algorithm introduces the trust information of users in a social network, but these algorithms usually use only the information of the adjacent user trust, ignoring the social network between users connected and indirect user trust influence, resulting in poor prediction accuracy.In view of this, the main work of this paper is divided into three parts, as follows:First of all, according to the social network connectivity we conduct iterative calculation to obtain social network user’s trust influence value; then, according to different user influence, adjacent similar users and trust influence factors were fused; finally, with the regularization, we use the product of users and items specific latent factors from the score matrix decomposition to predict user rating score on each item. In addition, the items similarity graph is introduced, and the characteristic quantity of the items in the scoring matrix is solved by the regularization method.Second:from view point of probability, our algorithm and the matrix decomposition based on items are deduced and integrated. The introduction of user trust network information can better solve the user cold start problem, at the same time, to solve item cold start, we introduce the item similar neighborhood information to the user ratings matrix and derive matrix decomposition based on the probability, and then get the gradient solution, and implemented the algorithm by LibRec. We find that through the contrast experiment, compared to the state-of-art social recommendation algorithms, this algorithm has better accuracy and can improve the recommendation system prediction performance for cold start users and items.Third:we implement a distributed version of the algorithm we proposed. Through the design of the model, we finally transform this score prediction problem to the problem of optimization. We conduct the comparison and analysis of the optimization method for solving such problems. Due to the high complexity of the algorithm, in order to deal with performance bottleneck for large data sets on stand-alone operation process, we design and implement a distributed algorithm based on Apache Spark especially for the iterative calculation.