Algorithms to Estimate Shapley Value Feature Attributions

Abstract

Black box machine learning models are increasingly prevalent. Their complex nature enables strong predictive accuracy but also makes them hard for humans to understand. One popular strategy to bridge the gap between complex models and interpretable models is to explain complex models using local feature attributions where a single sample's prediction is attributed to each of its features. In this class of explanation methods, Shapley value feature attributions have recently caught on. Although the Shapley value is appealing for its nice properties, it is NP-hard to compute in general. Here, we describe several works centered around tractably estimating the two most common variants of Shapley value feature attributions: marginal and conditional Shapley values.