GRALIS: Fusing Coalition and Gradient Attribution with Closed-Form Conservation Error and Finite-Sample Guarantees

GRALIS unifies competing explainability methods: The paper addresses a practical problem in responsible AI—that popular explanation methods like SHAP, LIME, GradCAM, and Integrated Gradients come from different theoretical foundations and often contradict each other. GRALIS combines the strengths of both "coalition-based" and "gradient-based" approaches into a single method, potentially reducing confusion about which tool to trust.

New mathematical guarantees improve reliability: Unlike existing methods, GRALIS provides two certified guarantees: (1) an exact calculation of how much explanation error occurs when attributing complex feature interactions (showing that order-d interactions are attributed at only 1/d of their true value), and (2) a finite-sample bound (O(1/√m) + O(1/k²)) that quantifies how accurate the method becomes with more data. These guarantees help practitioners understand the trustworthiness of explanations.

Theoretical foundation enables fair comparison: The paper proves that major attribution methods (SHAP, Integrated Gradients, LIME) can all be represented through a unified mathematical framework (the Riesz Representation Theorem). This theoretical bridge makes it possible to compare these previously incompatible methods fairly and understand their fundamental similarities and differences.

Addresses a real limitation in current practice: Prior research found that different explanation methods perform best on different faithfulness metrics with "method selection as the only proposed remedy." GRALIS offers an alternative by fusing complementary mechanisms rather than forcing practitioners to choose between imperfect options.