Explainability Methods for Deep Learning: A Comprehensive Benchmark and Practitioner Guide

The proliferation of explainability methods for deep neural networks has created confusion among practitioners seeking reliable and faithful explanations. We present the most comprehensive benchmark to date, evaluating 23 gradient-based, perturbation-based, and concept-based explainability methods across 12 quantitative metrics on 6 datasets spanning vision, NLP, and tabular domains. Our analysis reveals substantial inconsistencies: methods achieving high fidelity often perform poorly on sparsity, and gradient-based methods exhibit surprising robustness variations across architectures. We identify three clusters of explainability methods with distinct performance profiles and provide actionable recommendations for method selection based on application context. Our benchmark toolkit and evaluation framework are publicly released to standardize future comparisons.