Local Additive Feature Attribution: A Mathematical Taxonomy and Reporting Checklist

2026-07-15

AI Summary

This work addresses the lack of a unified framework in existing feature attribution methods, which leads to opaque assumptions, incomparable results, and susceptibility to failure modes. The authors propose the first unified mathematical framework for locally additive attributions, systematically integrating Shapley values, path integrals, gradient-based methods, perturbation approaches, and CAM-style techniques through five core dimensions: value functions, reference points, paths, perturbation distributions, and conservation rules. Through axiomatic analysis, comparative matrices, and formal modeling, the study elucidates how attribution outcomes depend critically on underlying assumptions and establishes causal links between methodological choices and characteristic failure modes. To enhance rigor, the paper concludes with a ten-item reporting checklist designed to substantially improve the transparency, reproducibility, and reliability of attribution research.

Abstract

Feature-attribution methods are central to explainable artificial intelligence. Their assumptions are expressed in several mathematical languages: cooperative-game values, path integrals, gradient operators, perturbation distributions, and backpropagation rules. This survey proposes a common framework for local additive feature attribution. It organizes Shapley, path-based, gradient/backpropagation, perturbation, and CAM-style methods around five specification choices: value function, reference, path, perturbation distribution, and conservation rule. It then compares these methods through an axiom-by-method matrix and links common failure modes, including baseline sensitivity, off-manifold perturbations, sanity-check failures, adversarial manipulation, and method disagreement, to the assumptions that produce them. Finally, the survey proposes a ten-item reporting checklist for studies that use local additive attributions. The central message is that attribution results are meaningful only relative to the mathematical assumptions under which they are defined, and that those assumptions should be reported.

Problem

• feature attribution
• explainable AI
• mathematical assumptions
• attribution methods
• failure modes

Innovation

• Local Additive Feature Attribution
• Mathematical Taxonomy
• Explainable AI
• Attribution Assumptions
• Reporting Checklist