Explainable Artificial Intelligence for Black-Box Machine Learning: A Taxonomy of Model-Agnostic and Deep Explanation Methods

Abstract

The rapid adoption of increasingly complex machine-learning models has intensified concerns regarding transparency, interpretability, accountability, and user trust. Deep neural networks, ensemble learners, and other high-capacity predictive models frequently provide superior predictive performance while offering limited insight into the mechanisms underlying individual predictions. Explainable artificial intelligence (XAI) has consequently emerged as an important research area concerned with making algorithmic decisions understandable to developers, domain specialists, regulators, and end users. This review examines the principal explainability approaches available through 2020 and organizes them according to model dependence, explanation scope, explanation representation, and underlying computational strategy. Particular attention is given to model-agnostic approaches such as Local Interpretable Model-Agnostic Explanations and SHapley Additive exPlanations, together with neural-network-specific approaches including saliency mapping, layer-wise relevance propagation, Grad-CAM, DeepLIFT, integrated gradients, and concept-based explanations. The review discusses the distinction between local and global interpretability, post-hoc and intrinsically interpretable models, and feature-attribution and example-based explanations. It further considers major limitations involving explanation fidelity, instability, computational cost, human interpretability, and the absence of standardized evaluation frameworks. The paper concludes that explainability should not be regarded as a single algorithmic property but as a multidimensional requirement involving the predictive model, explanation method, intended audience, and decision context.