Why-Not Questions & Explanations for Collaborative Filtering

Abstract

Throughout our digital lives, we are getting recommendations for about almost everything we do, buy or consume. However, it is often the case that recommenders cannot locate the best data items to suggest. To deal with this shortcoming, they provide explanations for the reasons specific items are suggested. In this work, we focus on explanations for items that do not appear in the recommendations they way we expect them to, expressed in why-not questions, to aid the system engineer improve the recommender. That is, instead of offering explanations on every item proposed by the system, we allow the developer give feedback about items that were not proposed. We consider here the most traditional category of recommenders, i.e., the collaborative filtering one, and propose ways for providing explanations for why-not questions. We provide a detailed taxonomy of why-not questions on recommenders, and model-specific explanations based on the inherent parameters of the recommender. Finally, we propose an algorithm for producing explanations for the proposed why-not questions.

Why-Not Explainable Graph Recommender

“…It is provided by means of instance-based (existing/missing input data points/tuples), query-based (faulty operators/data manipulations), or modification-based (modified queries/settings) explanations. In [32], the authors defined the problem of Why-Not questions for collaborative filtering RS and proposed an algorithm for producing Why-Not explanations tailored to the system developer. Recently, in [7] the authors consider Why-Not Questions in the context of top-k queries and score-based ranking functions.…”