Dummy Consistent Path Method for Robust Attribution with Irrelevant Features

Soyeon Kim, Seongwoo Lim, Kyowoon Lee, Jaesik Choi

Keywords: path method, input attribution, aumann shapley value, shapley interaction, dummy player axiom, dummy consistency axiom

TL;DR: We propose a path-based attribution method that resolves noisy explanations in rectified DNNs by dynamically finding a path that eliminates the influence of irrelevant (dummy) features, resulting in more reliable and consistent attributions.

Abstract: Attribution methods are widely used to interpret deep neural networks by identifying the input features that contribute to model predictions. Among these, path-based approaches, which accumulate gradients along a path from a baseline to an input, are often preferred for their theoretical foundations such as the Aumann-Shapley value. However, in networks with rectified activations, these methods frequently assign importance to irrelevant features—violating the Dummy Player axiom—due to the shattered gradients problem. Therefore, we propose the *Dummy Consistent Attribution Path Method* (DCAPM), a novel path-based approach designed to strictly enforce the exclusion of irrelevant feature interactions. Unlike standard approaches, our method dynamically constructs a path that nullifies spurious attributions arising from feature interactions with dummies to ensure zero attribution at every step, thereby ensuring theoretical consistency. To validate the practical impact of this theoretical adherence, we introduce a new metric that quantifies the alignment between dummy consistency and explanation faithfulness. We demonstrate that while existing methods systematically fail to satisfy this consistency, our approach yields significantly more robust and faithful explanations by adhering to the axiom.

Primary Area: interpretability and explainable AI

Submission Number: 12692