Research On Recommendation Algorithms And Recommendation Networks

Information filtering, as one of the technologies to solve the problem of information overload, has great importance in theory and practice. From the point of theorectical view, it is one of the issues in data mining. From the point of practical view, it is an important component in many e-commercial websites, which brings a lot of benefits for the on-line retailers. In recent years, information filtering has attracted many researchers from computer science, physical statistics, economics and mathematics. However, many problems are still not well solved, such as the diversity of recommendations as well as the personalized use of algorithms, sparsity of data sets and the characteristic of user-item bipartite networks. In order to solve those problems, we proposed some approaches using the theory and technique including computer science, physical statistics and mathematics:(1) Applied the multi-dimensional scaling(MDS) method in recommender systems to solve the diversity problem. MDS maps the high-dimensional data into low-dimensional data where the distances from the high dimension are well approximated by the distances from the low dimension. The method’s scalability is much better than SVD and Matrix factorization method. It is more efficient and accurate to compute similarities of item pairs by erasing the nosie in the data. The experimental results show that the accuracy and diversity of collaborative filtering method can be improved by 27.9% and 27.4%, respectively.(2) Proposed semi-local random walk method to solve the sparsity problem in recommender systems. Our method is based on the random walk method but taking into accout limited jumping steps. The accuracy and efficiency of our method are better than the random walk method and the coverage of items is much larger than other recommendation methods. The results show that our method is 10% better than other recommendation methods in sparse data. Our method ourperform other methods by 21.9% when recommending items for small degree users.(3) Proposed the social-based random walk algorithm. It takes into account not only the user-item bipartite network but also the user-group bipartite network. The random walk model on two networks can increase the reachable probability between nodes which improves the recommendation accuracy of small degree users. The results show that our method is 44.5% better than other methods when recommending items for small degree users.(4) Uncovered the information core in recommender systems. In this dissertation, we study the relevance of individual users and ?nd that there exists an information core whose size is small compared to the whole network. The users in the information core usually appear in many users’ top-N neighbor lists with high ranks. For many recommendation algorithms, one can achieve very good recommendation accuracy by only using the core users. The results show that our core user extraction method enables the recommender systems to achieve 90% of the accuracy of the top-L recommendation by taking only 20% of the users into account.(5) Proposed social-based matrix factorization algorithm. Based on matrix factorization method, we proposed a unified framework, which takes into account the rating, friendship and membership of users. Our model can solve the sparsity problem in recommender systems and provide diverse recommendations for users. The results show that the method’s accuracy can be improved by 19.5%、36.2% and 5% for item recommendation, group recommendation and friend recommendation, respectively.