Improving Jaccard index using genetic algorithms for collaborative filtering

Abstract

As data sparsity may produce unreliable recommendations in collaborative filtering-based recommender systems, it has been addressed by many researchers in related fields. Jaccard index is regarded as effective when combined with existing similarity measures to relieve data sparsity problem. However, the index only reflects how many items are co-rated by two users, without considering whether their ratings are evaluated similar or not. This paper proposes a novel improvement of Jaccard index, reflecting not only the ratio of co-rated items but also whether the ratings of each co-rated item by two users are both high, medium, or low. A genetic algorithm is employed to find the optimal weights of the levels of evaluations and the optimal boundaries between them. We conducted extensive experiments to find that the proposed index significantly outperforms Jaccard index on moderately sparse to dense datasets, in terms of both prediction and recommendation qualities.