Recommendation System Research Based On Latent Factor Model

With the rapid development of internet, internet of things, mobile internet etc, mass data leads to information overload problem. Especially in big movie recommendation sites, it's a hot and difficult topic that how to help users to get the information they want by a quickly and effective access. Among these researches, recommendation system is the main research object.Currently the alternative research directions at home and abroad include collaborative filtering recommendation, content-based recommendations, knowledge-based recommendation and hybrid recommendation. With the unique idea and convenient calculation, Collaborative Filtering Algorithm has been widely applied to personalized recommendation system. Followed with the increase of users and items, CF Algorithm has shortcomings on aspects of processing efficiency, sparse resistance and expanding.First the theory of latent factor model will be introduced. Then in order to solve the problem of sparse matrix that exist in the latent factor model, this paper will use the k-means algorithm to improve the latent factor model. So as to enhance the recommendation and forecast accuracy, we introduce the time context into the model. The main job of the article including: First of all, we introduce the principle and methods of recommendation systems. Second, this paper will introduce k-means algorithm to help solve the high sparse score matrix problem. It turned out that the k-means algorithm have an advantage in clustering. Based on the results of clustering, the score matrix will be reconstructed. Finally, by modifying the mark strategy in LFM, an improved new model K-LFM will be built. Third, in order to improve the precision of LFM model, this paper will add the time context information to the traditional user-item recommendation process. By changing the method of matrix decomposition, the three-dimensional matrix decomposition--“user-item-time” could be got to improve the model precision and model expanding. Forth, this paper will test the performance of several recommendation algorithms on movielens data set. What's more, a recommendation engine prototype will be implemented with the help of Mahout components on Hadoop platform.