Research On Collaborative Filtering Based Rating Prediction Algorithm

Web2.0 technology is leading the Internet into a new era. Users are the key part of the Internet. They both consume and produce data. The popularity of the Internet and the rapid growth of Internet users number have made an explosive growth of information, which leads to the “information overload” problem. Currently, the problem can be solved over two different kinds of technologies: Information Retrieval and Information Filtering.Search engine as a typical Information Retrieval technology has won great success. But search engines fail both to give user satisfied results when the users can’t give a precise description of the information he need and to push information to users actively. Recommender System, based on Information Filtering, has given an answer to this problem. It filter out the information that the user is interested in by exploiting users’ profile data and historical activities. Collaborative Filtering is the most broadly used algorithm in Recommender Systems. However, Collaborative Filtering is facing a series of challenges such as recommend accuracy or data sparseness. This paper mainly studies Collaborative Filtering. And we concentrate on model based especially matrix decomposition based and Restricted Boltzmann Machine based Collaborative Filtering algorithm. The contribution of this paper is mainly reflected in the following two aspects:First, we aim at finding the different impact of two key factors, neighbor number and similarity algorithm, on neighbor-based Collaborative Filtering in the view of accuracy. And at the same time, we compare the accuracy of different kinds of memory-based Collaborative Filtering algorithms based on the experiment on Movielens dataset.Second, we attempt to study and improve the Singular Value Decomposition based collaborative filtering algorithm-RSVD, and make use of the timing information, the user’s feature information and item’s feature information. We propose an improved algorithm-FeatureTRSVD, in which we integrate three useful biases: the time-context aware user bias and time-context aware item bias, the user register information bias and the item feature bias.Third, we give the improvement to the algorithm of Restricted Boltzmann Machine(RBM) for Collaborative Filtering, propose an item-based RBM approach for Collaborative Filtering in which we treat every item as an independent RBM, and the parameter is learned by the Batch Gradient Descent algorithm. The experimental results on the dataset of MovieLens show that the item-based RBM for CF outperforms user-based RBM for CF significantly and it also slightly outperforms Singular Value Decomposition method.