Counterfactual Explanations for Neural Recommenders

Abstract

While neural recommenders have become the state-of-the-art in recent years, the complexity of deep models still makes the generation of tangible explanations for end users a challenging problem. Existing methods are usually based on attention distributions over a variety of features, which are still questionable regarding their suitability as explanations, and rather unwieldy to grasp for an end user. Counterfactual explanations based on a small set of the user's own actions have been shown to be an acceptable solution to the tangibility problem. However, current work on such counterfactuals cannot be readily applied to neural models. In this work, we propose ACCENT, the first general framework for finding counterfactual explanations for neural recommenders. It extends recently-proposed influence functions for identifying training points most relevant to a recommendation, from a single to a pair of items, while deducing a counterfactual set in an iterative process. We use ACCENT to generate counterfactual explanations for two popular neural models, Neural Collaborative Filtering (NCF) and Relational Collaborative Filtering (RCF), and demonstrate its feasibility on a sample of the popular MovieLens 100K dataset. CCS CONCEPTS• Information systems → Recommender systems.

XRecSys: A framework for path reasoning quality in explainable recommendation

“…On one hand, post-hoc explainable models [18] consider recommendation and explanation generation as two distinct steps. Methods pertaining to this family either pre-compute paths in the knowledge graph and attach them to the recommended products generated by traditional models [5,1] or generate counterfactual explanations [14,13,9,7]. In both cases, the ranked products are optimized for utility and the training process of the recommendation model does not embed any constraint for the selection of accompanying reasoning paths.…”

Model-Agnostic and Diverse Explanations for Streaming Rumour Graphs

“…The reason being that rumours involve human perception, which is difficult to capture with discrete rules [7], whereas feature-based approaches fail to capture the graph-based structures of rumour propagation [8]. Counter-factual explanations, in turn, focus on features that may change the result of rumour detection rather than the features that led to the detection result in the first place [9].…”

“…Feature-based explanations, in turn, cannot capture the graphbased propagation structures of rumours [8]. Also, counterfactual approaches are not suited to understand why entities are classified as being part of a rumour, due to their focus on features that may change the result [9]. As such, explanations shall be based on a set of related examples, which enable users to generalize their properties [10].…”

“…The performance of existing modelspecific approaches are not generalized to other settings and even change considerably with a small of change of experimental setups [12]. Moreover, our approach is not limited to a single example as in many previous works [9], but include diverse examples to increase the likelihood of returning at least one option that matches the user's true interest. Especially, our approach is free of model bias, which can still happen in many modern ML models [13,14].…”

Explainable artificial intelligence (XAI) for healthcare decision-making

“…Wachter et al [117] modelled the process of identifying counterfactual examples as an optimization problem which ensures that the resultant examples have desired output and are close to the instance to be explained in the input space. Their idea has inspired subsequent research works [119,120]. Sharma et al [118] proposed a meta-heuristic evolutionary algorithm to identify counterfactual instances.…”