A new clustering method for collaborative filtering

Collaborative filtering, which predicts unknown ratings by analyzing the known ratings, is regarded as one of the most successful recommender systems within the last decade. In the process of collaborative filtering, clustering methods can recognize similar users and items. Then different methods can be applied to process different clusters. By these means, the accuracy and scalability of algorithms can be improved. Though the clustering methods for collaborative filtering have plenty of applications, they have long been neglected. The similarity measure for clustering was confused with the similarity measure for collaborative filtering. In this paper, we propose a new similarity measure for clustering and its application. Firstly, we use a basic similarity function to discover neighbor vectors of items. Secondly, we calculate the cosine similarity of the neighbor vectors for clustering. Thirdly, we finish the clustering process by using adjusted DBSCAN. For those users having many known ratings, we adjust the prediction function by adding a parameter which is the function of the item cluster size. The experiments on the grouplens dataset show that our method outperforms the previous methods and has been proved in many clustering applications in the field of collaborative filtering. For those users having many known ratings, our application method of clustering improves the prediction accuracy.