On the Impact of Jaccard Fusion with Numerical Measures for Collaborative Filtering Enhancement

Abstract

Collaborative filtering (CF) is an important method for recommendation systems, which are employed in many facets of our lives and are particularly prevalent in online-based commercial systems. The K-nearest neighbors (KNN) technique is a well-liked CF algorithm that uses similarity measurements to identify a user's closest neighbors in order to quantify the degree of dependency between the respective user and item pair. As a result, the CF approach is not only dependent on the choice of the similarity measure but also sensitive to it. However, some numerical measures, like cosine and Pearson, concentrate on the size of ratings, whereas Jaccard, one of the most frequently employed similarity measures, concerns the existence of ratings. Jaccard, in particular, is not a dominant measure, but it has long been demonstrated to be a key element in enhancing any measure. Therefore, in our ongoing search for the most effective similarity measures for CF, this research focuses on presenting combined similarity measures by fusing Jaccard with a multitude of numerical measures. Both existence and magnitude would benefit the combined measurements. Experimental results, on movielens-100K and Film Trust datasets, demonstrated that the combined measures are superior, surpassing all single measures across the considered assessment metrics.