Benefits of using symmetric loss in recommender systems

Abstract

The majority of online users do not engage actively with what they are offered: they mostly use few items, give feedback on even fewer. Additionally, in many cases, the only feedback available about the item is positive feedback. These issues are well-known in the area of personalized recommendation and there have been many attempts to develop recommendation algorithms based on data consisting of only positive feedback. Most such state-of-the-art recommendation methods use convex loss functions, and either interpret non-interactivity with an item as negative feedback or ignore such entries altogether, none of which in principal reflects the reality. In this work, we provide reasons to motivate the usage of a non-convex loss in implicit feedback scenario to deal with unlabelled data, and devise an algorithm to minimize the proposed loss in collaborative setting. We analyse the effects of the proposed loss both qualitatively and quantitatively on a benchmark public dataset.