The Research Of Personalized Recommendation System Based On Network Structure

The development of information and computer technology brings growing popularity of the Internet. Now, people are surrounded in the sea of all kinds of information and are facing with information overload. In this context, the recommendation system is brought forward which provides personalized service for different users. Throughout collecting and analyzing the users'information, the personalized recommendation system can predict the goods customers are possible interested in and recommend different goods for different customers so as to help the users find their favorite goods. On the other hand, The recommend system can effectively alleviate information overload, offer fast and convenient shopping experience and can attract more customers so as to improve the sales for electronic commerce. Generally, recommendation system is mainly composed of three parts: input module, recommendation module and output module. The recommendation module is the core of the system. A suitable and efficient recommendation method is a key to improve the performance of the system. As results, this thesis studies the bipartite network based on the user-item selections and evaluations relationship. We propose a recommending method with adaptively adjustment nodes weights. Then, by introducing the tag containing the information of tags'characters and categories, we propose a recommending method based the user-item-tag tripartite network. Using the actual users rating for the movie he has seen that the Movielen website provides, we test the performance of the proposed methods. The results shows that the proposed methods have good performance at recommending accuracy and diversity.The main contents of this thesis are as follows:First, this thesis systematically summarizes and analyzes the background and the current development of the personalized recommendation system. The current mainstream technologies are including: collaborative filtering techniques, content filtering and the methods based on the network structure. And then, we simply analyze the advantages and disadvantages of these three techniques. Second, we introduce the bipartite graph-based recommendation algorithm and describe the basic principle of bipartite graph, the system model and the specific implementation steps of the recommendation method. Furthermore, we summarize some improvement methods in literatures, compare the methods based on the bipartite network with other existing methods and analyze the problems in these methods and possible solutions.Third, The normalized rating that a user assigns to a item is weighted to the edge that connects the user and the item, which forms a user-item weighted bipartite-graph. We introduce an adjustment factor for items' degree in order to decrease the impact of the item with large degree on the items classification.In order to improve the system capacity of searching unpopular items,we introduce an adjustment factor for users' degree to enhance the impact of the users with large degree on the recommendation. Because they mainly select unpopular items. Throughout the tests on the Movielen datasets, we find the performance of the proposed algorithm is improved, compared with two standard recommendation algorithm based on the bipartite-graph.Fourth, we study the recommendation system model based on a user-item-tag tripartite graph. By introducing the items tag information, we calculate the diversity of the user interests. We increase the recommending weight that tags generate as the system recommends for a users having focusing interests, while increase the recommending weight that users generate as the system recommends for a user having wide interests. In this way, the system can make full use of advantages of these two kinds of information so as to gain better efficiency. Throughout the test by using the data of MovieLens, we find the proposed methods have good performance at the recommending accuracy and diversity.