Improved Collaborative Filtering Recommendation via Non-Commonly-Rated Items

Abstract

Collaborative filtering (CF) in recommendation systems has made great success in making automatic score predictions by using users' ratings on commonly-rated items. However, due to data sparsity and cold starting, in real systems, common-rated items among users are often not sufficient for accurate recommendations when using CF. Besides, the implicit relationships between users contained in huge amount of non-commonly-rated items are rarely utilized. In this paper, a new CF recommendation taking users' implicit relationships hidden in users' ratings on non-commonly rated items into consideration is proposed. In this method, we provide an algorithm to infer users' preferences for their non-commonly rated items and then based on these preferences. We obtain users' similarities on their non-commonly rated items. With a dynamic adjusting weight adapted to non-commonly rated items' proportion in two users' all rated items, we combine the similarities with traditional similarities based on co-rated items. Experiments are conducted on the MovieLens dataset for comparing the proposed approach with the traditional user-based collaborative filtering algorithm. The results show that our approach improves the recommendation accuracy.