Improving Jaccard index for measuring similarity in collaborative filtering

Abstract

In collaborative filtering-based recommender systems, items are recommended by consulting ratings of similar users. However, if the number of ratings to compute similarity is not sufficient, the system may produce unreliable recommendations. Since this data sparsity problem is critical in collaborative filtering, many researchers have made efforts to develop new similarity metrics taking care of this problem. Jaccard index has also been a useful tool when combined with existing similarity measures to handle data sparsity problem. This paper proposes a novel improvement of Jaccard index that reflects the frequency of ratings assigned by users as well as the number of items co-rated by users. Performance of the proposed index is evaluated through extensive experiments to find that the proposed significantly outperforms Jaccard index especially in a dense dataset and that its combination with a previous similarity measure is superior to existing measures in terms of both prediction and recommendation qualities.