Action-Triggering Recommenders: Uplift Optimization and Persuasive Explanation

A principal purpose of recommender systems is to induce positive user actions such as clicks and purchases. The performance of a recommender is typically evaluated in terms of prediction accuracy; a recommender is considered to be better if a larger number of its recommended items are purchased. However, accurate prediction is not enough for increasing user actions. The items might have been purchased even without recommendations, that is, the purchases are not triggered by the recommendations. Conversely, a user might be reluctant to purchase any items and a mere recommendation alone cannot motivate the user to purchase an item even if the recommended item is the best among all the items. In this work, we pursue recommender systems for triggering user actions. For this purpose, we tackle two issues: 1) uplift optimization and 2) persuasive explanation. For the first issue, we propose uplift-based evaluation and optimization methods for recommenders. Uplift, which is defined as an increase in user actions caused by recommendations, cannot be observed directly. We apply a causal inference framework to estimate the average uplift for the offline evaluation of recommenders. For optimization, we derive positive and negative samples specific to uplift and construct pointwise and pairwise optimization methods. Through experiments with three public datasets, we demonstrate the effectiveness of our optimization methods in improving uplift. For the second issue, we propose a new explanation style using context. The context style explanation presents contexts suitable for consuming the recommended items. The exhibited contexts make users imagine situations of items' usage and motivate them to purchase the items. Via a crowdsourcing-based user study, we confirm that the persuasiveness of our explanation style is better than conventional styles. The hybridization of context style with other styles further improves the persuasiveness.