Research And Application Of Random Walk Based Recommendation Techniques

With the rapid development of Internet, and information overload has become one important issue for web users to find useful information on web. In order to tackle these problems, researchers have developed personalized recommendation techniques. The work overviews related recommendation algorithms and focus on the application of random walk in personalized recommendation systems.In this work, we firstly introduce some classical neighborhood-based algorithms, which have been widely applied in business. Then, We analyze the sparsity of source data sets in recommendation system and the effects of sparsity in accuracy. Then, we overview previous related solutions, and introduce one random walk based method SimRank and its updated version SimRank++, which measures the similarity relationship between items with the assumption that the similarity is one relation with transitivity property. According to empirical results, we note that SimRank++produces poor performance when directly deployed to recommendation system. Thus, this paper works to improve this algorithm. In order to implement the personalized recommendation, we apply the topic-sensitive PageRank algorithm to the user-item bipartite graph, and obtain a new graph model: personalized PageRank (PPageRank). Also, sound theoretical analysis over the convergence of PPageRank algorithm is presented, and the relationship between PPageRank algorithm and graph model is bridged. Finally, the idea of SimRank++is employed to PPageRank, and results in the proposed Weighted PPageRank (WPPageRank).In this work, we select the user-based collaborative filtering (UserCF) and item-based collaborative filtering (ItemCF) algorithms as our baselines, and utilize the standard MovieLens data set for the evaluation experiments. The empirical results indicate that our proposed WPPageRank method and the improved SimRank++algorithm perform best.