Recommendation Methods Based On Network Analysis

With the rapid development of the Internet technology, information overload has become a serious problem Internet users facing. More and more papers and products information in academic sites and e-commerce sites make the users hard to quickly find their favorite papers or products. At the same time, how to accurately and real-timely provide the users interested papers and products is also the important issue what e-commerce enterprises are facing. Personalized recommendation system is an effective means to solve this problem.The author attended the work of constructing the Science "SocialScholar" academic network in Institute of Computing Technology of Chinese Academy. According to the needs of the project, we study the new methods for recommendation methods based on academic paper and design the corresponding paper on recommendation system. The main research work and results are as follows.(1) An algorithm for recommendation based on link prediction in a bipartite network is presented. We use a weighted bipartite network to represent the user-item matrix. Due to the similarity between recommendation and link prediction in bipartite network, we transform the recommendation into the problem of link prediction in bipartite network. The similarity based method is used to predict the potential links considering the topological similarity between the nodes in the network, as well as the similarity between users and the similarity between the items. The potential interests of the users to the items can be found by link prediction in the bipartite network. Our experimental results show that our algorithm can get high quality recommendation results.(2) We also propose a trust recommendation algorithm based on information dissemination. A weight combining the similarity and trust is assigned to each edge in the new trust network. A score vector of each node is transmitted in the trust network. Elements of the score vector are modified during the each iteration of transmission. Such transmission and iteration will be performed until convergence, and the score vector represents the final recommendation results. The algorithm can improve the performance in in terms of accuracy and diversity, and can alleviate the cold-start problems.(3) Finally, according to need of "SocialScholar" academic network construction, the paper recommendation system is built by applying the link prediction based recommendation algorithm and the recommendation algorithm based on trust and the dissemination of information. We analyze the needs, and propose the overall design of the target system framework, design the overall framework of the entire system, and embed the overall framework into multi-module hierarchy, achieve the recommendation features for users.