Perturbation-Based Explanations of Prediction Models

Abstract

Current research into algorithmic explanation methods for predictive models can be divided into two main approaches: gradient-based approaches limited to neural networks and more general perturbation-based approaches which can be used with arbitrary prediction...