Collaborative Filtering And Its Application In Recommender Systems

With the rapid development of Internet, the amount of information on the Web grows explosively, people often feel puzzled and helpless in finding and getting the intended information they really need. For overcoming this problem, recommender systems appeared and became a focus of researchers and practitioners. Recommender systems help users finding relevant information, products or services by providing personalized recommendations based on their profiles. Recommender systems are especially useful in an e-commerce environment, they enhance e-commerce sales in three ways:converting browsers into buyers; improving cross-sell by suggesting additional products for the user to purchase; improving loyalty by creating a value-added relationship between the site and the user.Collaborative filtering is one of the most successful and widely used techniques among recommender systems. It identifies users whose tastes are similar to those of the active user and recommends items that those users have liked. Collaborative filtering does not take into account content information, and is easier to implement. Most online shopping sites and many other applications now use the collaborative filtering technique to make personalized recommendations. However, despite its success and popularity, collaborative filtering suffers from a range of problems, such as sparsity, multiple-content, scalability and group recommendation. These problems limit the development of collaborative filtering, hence we should deeply study the problems.The main works of this dissertation are as follows:(1) The research on collaborative filtering at home and abroad is reviewed. Based on that, the basic knowledge of collaborative filtering technique is introduced, and the key issues in collaborative filtering are summarized.(2) The formation of neighborhood plays a fundamental role in collaborative filtering methods. Usually, similarities between an active user and other users are computed, and the k users with highest similarities are selected as active user’s neighbors. However, due to data sparsity, the accuracy of the similarity measures in collaborative filtering is decreased, which makes the formation of inaccurate neighborhood, thereby resulting in poor recommendations. To address this issue, this dissertation proposes a novel method of neighborhood formation: two-phase neighbor selection method. Definition of neighbor tendency is given. Based on the neighbor tendency, the preliminary neighborhood is formed. Then the equivalence relation similarity is applied to modify the preliminary neighborhood, which makes the formation of more accurate neighborhood. (3) To address the multiple-content problem in collaborative filtering, a collaborative filtering recommendation algorithm based on item-cluster preference is proposed. The proposed algorithm first finds out a set of candidate neighbors who are similar to the active user in item-cluster preference. The candidate neighbos have similar interest and more co-rated items with the active user. Then the algorithm identifies some nearest neighbors in the candidate neighbor set, which enhances the accuracy of searching for nearest neighbors.(4) To address the scalability problem in collaborative filtering, an improved affinity propagation algorithm is proposed. Grey relational analysis and Jaccard coefficient are applied to compute the user similarity matrix, which alleviates the data sparsity and enhances the clustering accuracy. For the n initial clusters produced by the algorithm, k clusters with the highest silhouette values are retained as basic clusters. Then the users who don’t belong to the basic clusters are assigned to the corresponding basic clusters, to generate the clustering results with the given number of clusters. Compared to the original affinity propagation algorithm, the proposed algorithm is efficient, it performs better in the application of addressing the scalability problem in collaborative filtering.(5) Although the traditional collaborative filtering recommender systems have achieved great success in recommending products to individuals, they are not suitable for group recommendation. As the number of groups increases rapidly in the virtual communities, building group recommender systems to provide personalized services to groups becomes more and more imperative. Therefore, a group recommendation algorithm combined with domain expert imputation is proposed. The proposed algorithm is designed based on the framework of item-based collaborative filtering. It first identifies group preferences according to every member’s preferences, and then generates recommendations based on the group preferences. Especially, the domain expert method is used to impute values for members’ unrated items in the recommendation process. In addition, the proposed algorithm considers the effects of member similarities on recommendation quality.(6) Based on the above theories and methods, a collaborative filtering recommender system, called myCFRS, is designed and developed. Main modules and functional structure of myCFRS are introduced, and development environment and system interfaces of myCFRS are given.