The Research Of Collaborative Filtering Recommendation Algorithm Based On Dynamic Statistics

With the rapid development of information technology and the Internet, users confront great amount of the information that with the exponential growth. It is more and more difficult and takes more time to find what they need actually for users with the rapidly growth of information. Recommended system comes out to solve these problems which users confront. It aims at providing personalized information filtering service according to the users’ preferences and interests. Currently, in a variety of recommendation technologies, collaborative filtering algorithm has outstanding performance and a lot of advantages compare to other recommended algorithms. Collaborative filtering has been used to varieties fields especially the great success in E-commerce system. However, Recommendation algorithm that based on collaborative filtering has a lot of shortcomings as follows: for examples data sparsity, cold start, system extensibility and so forth. More and more researchers have study on these problems.In the thesis, it introduce recommended system based on collaborative filtering, then analyzes of the various problems exist in collaborative filtering algorithm, and research how to improve the recommended system based on collaborative filtering. It proposes the calculation method of project popularity and user activity, then proposes the method to get none scoring matrix of users based on project popularity and the k nearest neighbors based on the user activity, and combine the two kinds of data to get the final result of recommended system. This thesis proposes the following studies:Firstly, in order to meet the needs of real-time systems, it needs us to collect the dynamic data of users and items. In this research, it proposes the calculation method of project popularity and calculation method across the dynamic data and so is the user activity. At last it reduces the user and item amount in this algorithm.Secondly, in order to meet the need of extensibility in recommended system, a sampling method is proposed based on the user’s activity. This method tries to reduce the algorithm complexity from n to a constant number. It greatly eases the problem which the time-consuming of recommended system increase rapidly when the users increase to a great number. So this method can improve cold start of collaborative filtering algorithm.Thirdly, since the cold start problem, this thesis proposes the fusion method of user’s data and items’ data. It calculates the none-rating scoring matrices according to rating scoring matrices, then proposes a compensation method for the none-rating scoring matrices based on item popularity. It can not only improve the real-time result in the system, but also solve the problem of new items. In the other way, the k-nearest neighbors based on the user activity and rating scoring matrices. Then the recommended system gets the result based on k-nearest neighbors and none-rating scoring matrices to improve the influence of new users. The proposed algorithm is greatly nake up for lack of collaborative filtering algorithm on a cold start according to these two steps.Finally, in order to verify the feasibility and effectiveness of the algorithm proposes in this thesis. Then compare and analyze the result between the proposed algorithm and traditional algorithms based on the dataset of Movie Lens. It proves that this algorithm reduce the running time significantly in ensuring the accuracy of the system.