An adaptive approach to collaborative filtering using attribute autocorrelation

Recommender Systems (RS) provide a rich collection of tools for enabling users to filter through large amount of information available on the Web. Collaborative Filtering (CF) is one of the most widely used and successful techniques behind the development of RS. CF based RS recommend items by computing similarities between users and/or items. The items recommended to a user are those preferred by similar users. However, with the tremendous growth in users and items on the Web, CF algorithms suffer from serious scalability problems because similarities between every pair of users and/or items need to be computed during the training phase. In this paper, we propose a scalable CF method by using data clustering techniques. The proposed work partitions the users of the CF system using an adaptive K-means clustering algorithm and then use those partitions (clusters) to select the similar users (neighborhood) of a target user. In this work, we also try to determine the optimal value of K (number of clusters). Once a target cluster is determined, the neighborhood of the target user is selected by looking into the similarity score between the target user and all other users in that cluster. The basic idea is to partition the users of the RS and apply the CF based recommendation algorithm separately to the partitions. The cluster-based approach reduces the runtime of the system as we avoid similarity computations over the entire rating data. Experiments performed on MovieLens-1M dataset indicate that our method is efficient in reducing the runtime as well as maintaining an acceptable recommendation quality.