Propagating Users' Similarity towards Improving Recommender Systems

Abstract

In this paper we examine an advanced collaborative filtering method that uses transitivity concepts. By propagating similarity between users, in a similar way as with we can significantly expand the space of potential recommenders and system's coverage, improving also the recommendations' accuracy. While information might be missing or be misleading and incorrect, similarity between two users can be directly calculated using the information from users' item ratings. A recent study observed a strong correlation between trust and preference in online rating systems, therefore it makes sense that transitivity concepts can also be applied to similarity, much as they are applied to trust. In contrast to a vast amount of work that seeks to exploit existed social information, like trust, from social networks to improve the recommendation process, we propose the other way round towards the same goal: use transitivity concepts exploiting the rating history of the recommender system's users to lead to the formation of new relationships and even social communities that were not previously existed. We propose a novel propagation scheme to confront the data sparsity problem in recommender systems and evaluate our method over two datasets with different characteristics, exhibiting a much higher recommendation coverage and better accuracy than classical collaborative filtering methods even under very sparse data conditions.