An evolutionary approach to interpretable learning

Abstract

Machine Learning (ML) interpretability is a growing field of computational research, of which the goal is to shine a light on black-box predictive models. We present an evolutionary framework to improve upon existing post-hoc interpretability metrics, by quantifying feature synergy, or the strength of feature interactions in high-dimensional prediction problems. In two problem instances from bioinformatics and climate science, we validate our results with existing domain research, to show that feature synergy is a valuable metric for post-hoc interpretability.