Interpreting Black-Box Models Across Domains with SHAP, LIME, and Counterfactuals

As black-box machine learning models—such as Random Forest and XGBoost ensemble frameworks—gain widespread adoption in high-stakes decision-making domains, the legal, moral, and technical demand for transparent and interpretable AI systems has grown immensely. This research presents a rigorous, multi-domain comparative evaluation of three prominent post-hoc Explainable AI (XAI) methodologies: SHAP (SHapley Additive exPlanations), LIME (Local Interpretable Model-agnostic Explanations), and Counterfactual Explanations (via DiCE). Using diverse, real-world tabular datasets representing the healthcare, finance, and demographic domains, this study benchmarks the selected XAI methods across three critical evaluation axes: explanation stability under localized input perturbations, human-centric practical interpretability, and computational runtime efficiency. The empirical experimental results reveal clear performance trade-offs among the evaluation paradigms: SHAP delivers highly faithful, theoretically robust, and consistent global and local feature attributions, albeit at a higher computational runtime expense. LIME serves as a highly efficient, fast alternative for real-time local instance-level explanations, though it exhibits greater sensitivity and fluctuation under subtle feature perturbations. Counterfactual Explanations (DiCE) complement feature-attribution strategies by providing actionable, human-comprehensible "what-if" scenarios that are highly valuable in clinical and financial decision-justification pathways. Ultimately, this paper bridges the gap between theoretical interpretability research and field application, introducing a standardized benchmarking infrastructure and practical, domain-specific recommendations to assist data science professionals in deploying reliable, accountable, and transparent machine learning workflows.