Pairwise loss regularization for recommendations explanation
Abstract
Recommender systems are notoriously complex systems for which providing a local explanation on why a certain item is proposed to a specific user is still a challenging task. Most explanation approaches focus on predicting the rating of items thereby minimizing some discrepancy with the real ratings by means of traditional loss functions (e.g., sum of squares error). However, most of the times, ratings may not fully embrace user preferences concerning the ranking of items. To better embrace user preferences, methods based on ranking losses have been proposed either to recommend or to explain why an item is recommended. Although effective at identifying the most prominent items, these methods fail to capture a realistic value for the rating attached to their explanation. This loss attached to the semantic of the recommendation can in turn arm the trust of a user in the explanation. In this paper, we propose and discuss experimental results of a simple yet effective novel loss schema that balances ranking and rating losses to provide a best of both world explanation.
Domains
Computer Science [cs]
Origin : Files produced by the author(s)
licence : CC BY NC - Attribution - NonCommercial
licence : CC BY NC - Attribution - NonCommercial