Let Me Explain, Again: Multiplicity in Local Sufficient Explanations

TMLR Paper6784 Authors

Abstract: When asked to explain their decisions, humans can produce multiple complementary justifications. In contrast, several feature attribution methods for machine learning produce only one such attribution, despite the existence of multiple equally strong and succinct explanations. The explanations found by these methods thus offer an incomplete picture of model behavior. In this paper, we study the problem of explaining a machine learning model's prediction on a given input from the perspective of minimal feature subsets that are sufficient for the model's prediction, focusing on their non-uniqueness. We give a tour of perspectives on this non-uniqueness, in terms of Boolean logic, conditional independence, approximate sufficiency, and degenerate conditional feature distributions. To cope with the multiplicity of these explanations, we propose a wrapper methodology that can adapt and extend methods that find a single explanation into methods for finding multiple explanations of similar quality. Our experiments benchmark the proposed meta-algorithm, which we call Let Me Explain Again (LMEA), against two multi-explanation method baselines on synthetic and real-world multiple-instance learning problems for image classification and demonstrate the ability of LMEA to augment two single-explanation methods.

Submission Type: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Dennis_Wei1

Submission Number: 6784