| With the rapid development of Web 2.0 technology, the internet activities have been an impartible part in our life. At the meantime, with the pushing of variety mobile smart terminal technique, the information data on the social network present explosive growth. In such a situation, how to exactly choose the information that the users really need and mine the useful information from the huge amount of data has the important theoretical value and practical significance. Generally speaking, people often need to make forecasts about the information's future or the development tendency according to the information obtained from the huge data which has been known. Link prediction and its application have been appeared at the time. Link prediction has been applied on many situation, and one of the most popular application is recommendation system. As we know, the traditional recommendation algorithms are almost based on the vertex similarity. However, most information of the nodes on the actual social network is hidden. So the traditional recommendation algorithms couldn't satisfy the demands on the social network. Therefore, how to recommend accuracy information to people from the huge data on the social network has become a really significant problem.In this paper, we study the relevant theories of social network and analyze some algorithms of social network analysis, then we find out Random Walk is widely used on social network analysis, especially on link prediction and recommendation system. As the random walk is simple and easy to understand, many researchers have made improvements on it, we also try to make some relevant improvements based on the social network's attributes. We first add topological attribute to Random Walk Model, and obtain a nice result on the link prediction,then we try to apply the improved Random Walk on the recommendation system. As it mentioned before, there are variety hidden attributes in the social network which may influence the recommendation system, we exploit the attributes in the social network to improve the Random Walks so that the accuracy of recommendation can be improved. As we studied, most of the attributes are text information, and Latent Dirichlet Allocation(LDA) theme model has done a nice job on mining potential information from text. So we firstly use LDA model to mine potential themes of the nodes in the social network and regard the themes as the potential attributes of the nodes, then combing with the information that already obtained so that we can improve recommendation system's performance much better.Link prediction in complex networks has been an attractive problem. Generally, when we obtain a snapshot of a network, we would like to infer which interactions are most likely to occur among the existing members in the future. This kind of problem has been extensively studied both from academia and industry. However, link prediction is challenging in practice. To address this issue, in this paper, we focus on the topology structure of the networks for link prediction. Two algorithms CN-LRW and CN-RWR are proposed based on local random walk and random walk with restart, respectively. We evaluate our approaches on three real data sets. Experiments show that CN-LRW and CN-RWR outperform LRW and RWR respectively in most cases. Incorporating node information with the network structure features seems promising in discovering the latent semantics of the network.Recommendation systems become extremely popular and widely applied in recent years. Researchers have done much work to developing recommender systems in social network. However, most of the methods only recommend relationships or products separately. To address this problem, we propose an User-Item(UI) bipartite graph which simultaneously incorporates relationships and interest information to model complex relationships among users and products. And then we use LDA to explore the potential relationship of the products, finally we apply Random Walk on the UI bipartite graph to measure the relevance between users and products, which may recommend products to users. We evaluate our approach on Cite ULike dataset and last.fm dataset. Experiments show the effectiveness of our approach. Comparison with other methods on the two datasets indicates that our approach do make a better job. Besides, considering the high computation complexity, we apply a biased absorbing random walk on the bipartite graph to reduce the computation complexity. We further add topology and potential attributes to the absorbing random walks, so that we can analyze the influence of potential links on the recommendation system. The results show the proposed algorithms TB-ARW and LA-ARW have a better performance on items recommendation system. |
PDF Full Text Request