Learning Counterfactual Explanations for Recommender Systems

Oren Barkan, Veronika Bogina, Liya Gurevitch, Yuval Asher, Noam Koenigstein

Keywords: Recommender Systems, Explanations

TL;DR: A novel framework for extracting counterfactual explanations for recommender systems.

Abstract: We introduce the Learning to eXplain Recommendations (LXR) framework, a post-hoc, model-agnostic framework for counterfactual explanations of recommender systems. LXR can work with any differentiable recommender and learns to score the importance of users' personal data with respect to a recommended item. The framework's objective employs a novel self-supervised counterfactual loss term that seeks to spotlight the user data most instrumental in the recommendation of an item. Additionally, we propose several counterfactual evaluation metrics for assessing explanations in recommender systems. Using these metrics, our results demonstrate LXR's capability to provide counterfactual explanations for various recommendation algorithms across different datasets. LXR's code is publicly available at https://github.com/ExplainingRecommendations/LXR.

Track: User Modeling and Recommendation

Submission Guidelines Scope: Yes

Submission Guidelines Blind: Yes

Submission Guidelines Format: Yes

Submission Guidelines Limit: Yes

Submission Guidelines Authorship: Yes

Student Author: No

Submission Number: 1449