Preference-based and local post-hoc explanations for recommender systems. Issue 108 (September 2022)

• Record Type: Journal Article
• Authors: Brunot, Léo; Canovas, Nicolas; Chanson, Alexandre; Labroche, Nicolas; Verdeaux, Willème
• Is Part Of: Information systems. Issue 108(2022)
• Publication Date: 2022-09
• Subjects: Post-hoc explanation -- Recommender systems -- Locality -- Pairwise preference; Database management -- Periodicals; Electronic data processing -- Periodicals; Bases de données -- Gestion -- Périodiques; Informatique -- Périodiques; Database management; Electronic data processing; Periodicals; 005.7
• DOI: 10.1016/j.is.2022.102021
• Languages: English
• ISSNs: 0306-4379

Abstract

Abstract: Post-hoc explanation aims at defining a simple local surrogate model to shed light on a prediction produced by a complex, generally black-box, model. In the general context of classification, it has been shown that local surrogates may not always be able to capture a local explanation, i.e. for a specific instance prediction, but rather depict more of a general behavior of the black-box. This problem is even more complex in a recommendation scenario where classes and decision boundaries are not explicitly defined and where data are very sparse by nature. We show in this paper that it is possible to tackle these problems with an efficient sampling around the recommendation instance to explain, to finally learn a proper local surrogate model. To this aim, this paper introduces several new approaches to capture efficiently local explanation models in the context of recommendation, all defined around a locality sample. Noticeably, and novel to this work, we show that it is possible to achieve a simple, yet better quality explanation model by not directly considering ratings, but rather implicit preferences as expressed by comparisons of pairs of ratings. We introduce to this extent a novel explainable model based on a pairwise loss RankNet architecture. Extensive experiments show that our methods can be better than state-of-the-art methods depending on the locality of the black-box model, and are much more efficient to retrieve meaningful explainable features locally. Highlights: Formal definitions of an explanation instance and interpretable features. In-depth study of locality and perturbation mechanism around an explanation instance. Model agnostic post-hoc linear explanation model with pointwise loss. Extension of linear-based explanation based on pairwise preference loss. Extensive experiments with 4 recommenders and up to 20M ratings.