Faithful Post-hoc Explanation of Recommendation Using Optimally Selected Features

Abstract

Recommendation systems have improved the accuracy of recommendations through the use of complex algorithms; however, users struggle to understand why the items are recommended and hence become anxious. Therefore, it is crucial to explain the reason for the recommended items to provide transparency and improve user satisfaction. Recent studies have adopted local interpretable model-agnostic explanations (LIME) as an interpretation model by treating the recommendation model as a black box; this approach is called a post-hoc approach. In this chapter, we propose a new method based on LIME to improve the model fidelity, i.e., the recall of the interpretation model to the recommendation model. Our idea is to select an optimal number of explainable features in the interpretation model instead of using complete features because the interpretation model becomes difficult to learn when the number of features increases. In addition, we propose a method to generate user-friendly explanations for users based on the features extracted by LIME. To the best of our knowledge, this study is the first one to provide a post-hoc explanation with subjective experiments involving users to confirm the effectiveness of the method. The experimental evaluation shows that our method outperforms the state-of-the-art method, named LIME-RS, with a 2.5%–2.7% higher model fidelity of top 50 recommended items. Furthermore, subjective evaluations conducted on 50 users for the generated explanations demonstrate that the proposed method is statistically superior to the baselines in terms of transparency, trust, and satisfaction.