New Taxonomy and Checklist for Explaining AI Feature Attributions

Feature attribution methods are fundamental to explainable AI (XAI), helping users understand which input features contribute most to a model's prediction. However, these methods, such as Shapley values, path integrals, and gradient-based approaches, are often described using diverse mathematical languages, leading to confusion and inconsistent application. This survey introduces a unified mathematical framework for local additive feature attribution. It systematically categorizes these methods based on five key specification choices: the value function, reference point, path, perturbation distribution, and conservation rule. By doing so, it clarifies the underlying assumptions of each method and helps to demystify their operational principles. Furthermore, the paper links common failure modes of XAI methods—like baseline sensitivity, off-manifold perturbations, and adversarial manipulation—directly to these specific mathematical assumptions. To promote better practice and transparency, the survey concludes with a ten-item reporting checklist for researchers and practitioners using local additive attributions, emphasizing that attribution results are only meaningful when their defining assumptions are clearly stated.