Abductive Explanations for Groups of Similar Samples

Daniel Díaz Quílez, Tuomo Lehtonen

Keywords: explainable AI, robust explanations, group explanations, abductive explanations, neural network verification

TL;DR: We introduce $\delta$–robust abductive explanations, which allow for producing feature selection explanations valid for a group of similar samples

Abstract: Explaining the decisions of machine learning models is crucial as their use becomes widespread. While many approaches to explanation are based on heuristics or surrogate models without formal guarantees, formal explanations provide reasoning for a particular decision that is guaranteed to be valid. We focus on abductive explanations (AXp) that identify sufficient subsets of input features for a given classification. We extend AXp to not only cover a particular sample, but to cover all of the samples whose features are within a given interval, providing explanations that remain valid even when the features in the explanation vary by up to $\delta$. In addition to applying this notion of $\delta$-robust AXp to a single sample, we also consider \emph{group explanations} ($\delta$-gAXp), which give a common explanation for a group of samples that share the same classification. We evaluate our approach by producing explanations for neural networks with the help of Marabou, a neural network verifier. The evaluation shows that, compared to a recent approach for finding a maximally ``inflated'' explanation, a $\delta$-robust AXp covers a significant volume of the inflated explanation with a dramatically lower runtime. Our evaluation also provides evidence that group explanations capture important features for all the samples within the group much faster than computing explanations for each sample separately.

Primary Area: interpretability and explainable AI

Submission Number: 24541