Local Interpretable Model-agnostic Explanations of Bayesian Predictive Models via Kullback-Leibler Projections

TL;DR

KL-LIME is a novel method that explains Bayesian model predictions locally by projecting complex predictive distributions onto simpler interpretable models, balancing fidelity and complexity.

Contribution

It combines LIME with Bayesian projection techniques to improve local explanations of Bayesian predictive models.

Findings

Effective explanation of Bayesian neural network predictions on MNIST.

Balances explanation fidelity and interpretability using information theory.

Demonstrates applicability to deep convolutional neural networks.

Abstract

We introduce a method, KL-LIME, for explaining predictions of Bayesian predictive models by projecting the information in the predictive distribution locally to a simpler, interpretable explanation model. The proposed approach combines the recent Local Interpretable Model-agnostic Explanations (LIME) method with ideas from Bayesian projection predictive variable selection methods. The information theoretic basis helps in navigating the trade-off between explanation fidelity and complexity. We demonstrate the method in explaining MNIST digit classifications made by a Bayesian deep convolutional neural network.