Model Agnostic Multilevel Explanations

TL;DR

This paper introduces a model-agnostic multilevel explanation tree that enhances interpretability of black-box models by providing local, group, and global explanations, improving communication and efficiency in understanding model decisions.

Contribution

It proposes a novel meta-method to construct multilevel explanation trees from local explainability methods, incorporating clustering and side information for better interpretability.

Findings

Effective in providing explanations at multiple levels

Validated through human studies with experts and non-experts

Produces high-fidelity sparse explanations on real datasets

Abstract

In recent years, post-hoc local instance-level and global dataset-level explainability of black-box models has received a lot of attention. Much less attention has been given to obtaining insights at intermediate or group levels, which is a need outlined in recent works that study the challenges in realizing the guidelines in the General Data Protection Regulation (GDPR). In this paper, we propose a meta-method that, given a typical local explainability method, can build a multilevel explanation tree. The leaves of this tree correspond to the local explanations, the root corresponds to the global explanation, and intermediate levels correspond to explanations for groups of data points that it automatically clusters. The method can also leverage side information, where users can specify points for which they may want the explanations to be similar. We argue that such a multilevel structure can also be an effective form of communication, where one could obtain few explanations that characterize the entire dataset by considering an appropriate level in our explanation tree. Explanations for novel test points can be cost-efficiently obtained by associating them with the closest training points. When the local explainability technique is generalized additive (viz. LIME, GAMs), we develop a fast approximate algorithm for building the multilevel tree and study its convergence behavior. We validate the effectiveness of the proposed technique based on two human studies -- one with experts and the other with non-expert users -- on real world datasets, and show that we produce high fidelity sparse explanations on several other public datasets.