A Unified Taxonomy of Explainability: Theoretical Foundations for XAI Methods

Abstract

The increasing integration of AI into high-stakes sectors such as healthcare, finance, autonomous systems, and criminal justice has heightened the need for transparent, interpretable decision-making processes. A key research area is Explainable Artificial Intelligence (XAI), but the field struggles with unclear concepts, varied language, and scattered categorization. This paper introduces a comprehensive theoretical framework for classifying XAI methods by combining existing taxonomies, laying down formal underpinnings, and addressing new challenges. In this paper, a multi-dimensional taxonomy is proposed, organized into five main areas: temporal integration (before vs. after an event), model dependency (specific to a model vs. general), explanation scope (narrow vs. broad), transparency level (built-in vs. added later), and explanation granularity (features, concepts, examples, or rules). The framework clarifies terminological confusion by drawing formal distinctions among interpretability, explainability, transparency, and understandability. Furthermore, the paper presents a sixth dimension, regulatory compliance (e.g., GDPR’s right to explanation, EU AI Act’s transparency requirements), and illustrates how legal obligations define taxonomic boundaries. This research enhances the taxonomy by incorporating Large Language Model (LLM) explanation methods, a new domain absent from previous classifications. To systematically assess the quality of XAI methods across taxonomic categories, this paper defines four evaluation metrics: fidelity, stability, comprehensibility, and completeness. Compared with five previous frameworks, the new taxonomy offers superior discrimination and is backward-compatible. This research provides a strong theoretical basis that integrates current techniques, new approaches, and regulatory requirements into a cohesive, adaptable classification framework.

Ähnliche Arbeiten