Local Additive Feature Attribution: A Mathematical Taxonomy and Reporting Checklist
This paper introduces a common mathematical framework for local additive feature attribution methods in explainable AI, categorizing techniques and linking their failure modes to underlying assumptions. It also proposes a ten-item reporting checklist to enhance transparen…
Intelligence analysis by Gemini 2.5 Flash
A new survey paper provides a unified mathematical taxonomy for diverse local additive feature attribution methods, which are crucial for explaining AI decisions. It organizes existing techniques based on five specification choices, compares them through an axiom-by-method matrix, and identifies how common failure modes stem from specific assumptions. The authors conclude by offering …
Imagine you have a robot that makes decisions, but you don't know why. Scientists use special tools to figure out which parts of the information the robot saw were most important for its decision. This paper is like a guide that helps scientists understand all the different kinds of these tools, how they work, and why some might give confusing answers. It also gives them a checklist to make sure they explain their findings clearly, so everyone can trust the robot's explanations.
Analysis
A Unified Framework for XAI Attribution
Feature attribution methods are fundamental to explainable artificial intelligence (XAI), providing insights into why an AI model makes a particular decision. However, these methods often operate under diverse mathematical assumptions, expressed through various languages like cooperative-game values, path integrals, and gradient operators. This paper addresses this fragmentation by proposing a common framework for local additive feature attribution, aiming to unify the understanding of these disparate techniques. It organizes methods such as Shapley, path-based, gradient/backpropagation, perturbation, and CAM-style approaches around five core specification choices: the value function, reference, path, perturbation distribution, and conservation rule. This structured approach allows researchers to systematically compare and contrast different attribution methods, highlighting their underlying mechanisms and design decisions.
Unpacking Method Assumptions and Failure Modes
The survey meticulously compares these diverse methods using an axiom-by-method matrix, which serves as a powerful tool for understanding their theoretical underpinnings. By mapping methods against a set of axioms, the paper reveals the specific assumptions each technique relies upon. Crucially, it links common failure modes observed in XAI—such as baseline sensitivity, off-manifold perturbations, sanity-check failures, adversarial manipulation, and method disagreement—directly to these underlying mathematical assumptions. This connection is critical for diagnosing why certain attribution methods might produce unreliable or misleading explanations in specific contexts, providing a deeper understanding of their limitations and appropriate use cases.
Towards Transparent Reporting in XAI
Recognizing that the meaningfulness of attribution results is inherently tied to the assumptions under which they are defined, the paper culminates in the proposal of a ten-item reporting checklist. This checklist is designed to guide studies that utilize local additive attributions, ensuring that researchers explicitly report the mathematical assumptions and choices made when applying these methods. The central message is clear: for attribution results to be truly interpretable and trustworthy, the specific mathematical context and assumptions must be transparently communicated. Adopting this checklist could significantly improve the rigor, reproducibility, and interpretability of XAI research, fostering greater confidence in the explanations provided by AI systems.
Key points
- The paper proposes a common mathematical framework for local additive feature attribution methods in explainable AI.
- It organizes diverse attribution techniques (Shapley, path-based, gradient, perturbation, CAM-style) around five specification choices.
- An axiom-by-method matrix is used to compare methods and link common failure modes to their underlying mathematical assumptions.
- The survey introduces a ten-item reporting checklist to improve transparency and reliability in studies using local additive attributions.
- The central message emphasizes that attribution results are meaningful only when their defining mathematical assumptions are clearly reported.
If researchers widely adopt this proposed framework and reporting checklist, it could lead to more consistent, reliable, and trustworthy explanations from AI systems. This standardization would foster greater confidence in AI applications across various domains and accelerate research by providing a common language for comparing different attribution methods.
Without widespread adoption of the proposed checklist and unified framework, the existing issues of method disagreement, baseline sensitivity, and other failure modes in XAI could persist. This might lead to continued misinterpretation of AI explanations, potentially undermining trust and hindering the responsible and ethical development of AI.
