Read between the interactions: Understanding non-interacted items for accurate multimedia recommendation

Abstract

This paper addresses the problem of multimedia recommendation that additionally utilizes multimedia data, such as visual and textual modalities of items along with the user-item interaction information. Existing multimedia recommender systems assume that all the non-interacted items of a user have the same degree of negativity, thus regarding them as candidates for negative samples when training the model. However, this paper claims that a user?s non-interacted items do not have the same degree of negativity. We classify these non-interacted items of a user into two kinds of items with different characteristics: unknown and uninteresting items. Then, we propose a novel negative sampling technique that only considers the uninteresting items (i.e., rather than the unknown items) as candidates for negative samples. In addition, we show that using the multiple Bayesian personalized ranking (BPR) losses with both unknown and uninteresting items (i.e., all the non20 interacted items) in existing multimedia recommendation methods is very effective in improving recommendation accuracy. By conducting extensive experiments with three real-world datasets, we show the superiority of our ideas. Our ideas can be easily and orthogonally applied to any multimedia recommender systems.