Skip to Main Content
Collaborative filtering, one of the most widely used approach in recommender system, predicts a user's rating towards an item by aggregating ratings given by users having similar preference to that user. In existing approaches, user similarity is often computed on the whole set of items. However, because the number of items is often very large, and so is the diversity among items, users who have similar preference in one category of items may have totally different judgement on items of another kind. In order to deal with this problem, we propose a method of clustering items, so that inside a cluster, similarity between users does not change significantly. After that, when predicting rating of a user towards an item, we only aggregate ratings of users who have high similarity degree with that user inside the cluster to which that item belongs. Experiments evaluating our approach are carried out on the real dataset taken from movies recommendation system of MovieLens web site. Preliminary results suggest that our approach can improve prediction accuracy compared to existing approaches.