ARCHEX: GPU-Accelerated Archetypal Explanation for Discovering Complementary Patterns in Model Interpretability

While local explainability methods like SHAP and LIME have become standard tools for model interpretation, they often miss the global structural patterns that define data archetypes. We introduce \textbf{ARCHEX} (ARCHetype-based EXplainer), a GPU-accelerated sparse archetypal analysis framework that identifies interpretable extreme data profiles. Through comprehensive experiments on five diverse datasets, we demonstrate that ARCHEX discovers fundamentally different patterns than local methods (Spearman $\rho$ as low as 0.075), providing unique insights into model behavior. Our adaptive $k$-selection algorithm automatically determines optimal archetype complexity, while bootstrap confidence intervals ensure statistical rigor. On the Breast Cancer Wisconsin dataset, ARCHEX identifies fractal dimension as the most important feature (ranked \#19 by SHAP), revealing morphological patterns invisible to local methods. Our GPU implementation achieves 10--100$\times$ speedup over CPU baselines, enabling real-time interpretation. This work establishes ARCHEX as a powerful complement to existing XAI techniques, particularly valuable when global data structure matters.