Survey on Explainable AI for Traditional Machine Learning and Domains

Abstract

The increasing deployment of opaque AI models in high-stakes domains has intensified the demand for Explainable AI (XAI) that is both cognitively aligned and operationally embedded. This survey reconceptualizes explainability as a reflexive, system-level property spanning the entire machine learning lifecycle. We introduce two novel dimensions: (i) a cognitively grounded taxonomy of explanation strategies–including analogical, contrastive, conceptual, narrative, and interactive forms–aligned with human reasoning models; and (ii) a lifecycle-centric architecture that embeds explainability across four interdependent layers: Operational, Explainability, Interactivity, and Governance. Through a systematic review of 202 peer-reviewed studies, we analyze trends in explanation formats, evaluation metrics, and domain-specific adaptations. We further present a comprehensive benchmark of 17 XAI techniques across tabular, image, and text modalities, evaluated using lifecycle-aware and cognitively aligned metrics such as fidelity, completeness, monotonicity, stability, and complexity. Together, these contributions offer a unified foundation for designing, evaluating, and deploying transparent, trustworthy, and human-centered AI systems.

Ähnliche Arbeiten