Authors: Vaibhav Singh Chouhan, Rupali Chaure
Abstract: Image caption: Word Cloud for Explainable A Explainable Artificial Intelligence (XAI) is an important subject area… This extensive review paper provides a meta-analysis of the most prominent explainable artificial intelligence (XAI) frameworks, methodologies and techniques that have been developed in the last decade. We conduct a systematic analysis of > 150 peer-reviewed publications to integrate novel transparent techniques (attention mechanisms, LIME, SHAP, prototype-based methods, counterfactual explanations) and assess their usefulness for promoting interpretability and user trust. Our meta-analysis exposes severe shortcomings in interpretability standardization, validation metrics and real-world applicability. We introduce a universal taxonomy to classify XAI methods based on their explanation scopes, computational complexity and applicability for different application domains. In addition we discuss the accountability-interpretability tradeoff, scalability issues and the need for domain specific explanation frameworks as key challenges still facing this field. Our paper contributes to the field by offering a holistic roadmap that will guide researchers and practitioners to select, implement and evaluate an XAI solution, trace future research paths which are required in order to endow autonomous decision-making systems with trustworthiness when applied on critical infrastructure.
DOI: http://doi.org/10.5281/zenodo.21545756
