A survey of feature attribution techniques in explainable AI: taxonomy, analysis and comparison

Abstract

The feature attribution methods have become central to explainable artificial intelligence (XAI), providing critical insights into how machine learning (ML) models make individual and aggregate decisions. This survey presents a complete taxonomy of feature attribution techniques, organizing them into model-agnostic and model-specific categories while highlighting extensions such as rule-based and attention-based explanations. We analyze formal definitions of each method, mathematical formulations, application contexts, strengths, and limitations. A comparative analysis highlights key trade-offs among model flexibility, computational cost, explanation fidelity, and interpretability. In addition to theoretical perspectives, we provide practical comparisons of selected methods on benchmark tasks to guide real-world applicability. The emerging trends toward global interpretability, hybrid attribution approaches, and human-centered evaluation frameworks are discussed. This survey synthesizes current advancements and presents future directions for developing scalable, robust, and user-aligned feature attribution methods to advance responsible and transparent AI.