Active Learning Applied to Rating Elicitation for Incentive Purposes

Abstract

Active Learning (AL) has been applied to Recommender Systems so as to elicit ratings from new users, namely Rating Elicitation for Cold Start Purposes. In most e-commerce systems, it is common to have the purchase information, but not the preference information, i.e., users rarely evaluate the items they purchased. In order to acquire these ratings, the e-commerce usually sends annoying notifications asking users to evaluate their purchases. The system assumes that every rating has the same impact on its overall performance and, therefore, every rating is worth the same effort to acquire. However, this might not be true and, in that case, some ratings are worth more effort than others. For instance, if the e-commerce knew beforehand which ratings will result in the greatest improvement of the overall system’s performance, it would be probably willing to reward users in exchange for these ratings. In other words, rating elicitation can go together with incentive mechanisms, namely Rating Elicitation for Incentive Purposes. Like in cold start cases, AL strategies could be easily applied to Rating Elicitation for Incentive Purposes in order to select items for evaluation. Therefore, in this work, we conduct a extensive benchmark, concerning incentives, with the main AL strategies in the literature, comparing them with respect to the overall system’s performance (MAE). Furthermore, we propose a novel AL strategy that creates a k-dimensional vector space, called item space, and selects items according to the density in this space. The density-based strategy has outperformed all others while making weak assumptions about the data set, which indicates that it can be an efficient default strategy for real applications.