Exact Shapley Attributions in Quadratic-time for FANOVA Gaussian Processes

Shapley values provide principled feature importance explanations, but have been computationally impractical: The paper addresses a major limitation of Shapley values—a theoretically sound method for explaining AI decisions—which previously required exponential computation time, making them unusable for real-world applications with many features.

This research enables exact explanations for a specific but important class of AI models: For FANOVA Gaussian Processes (a type of probabilistic model), the authors achieved quadratic-time computation of exact Shapley values, meaning explanations that were previously impossible to calculate can now be generated efficiently for this model class.

The method handles uncertainty in AI predictions, not just point estimates: Unlike simpler explanation methods, this approach works with probabilistic models that output ranges of possible predictions rather than single answers, capturing both what the model predicts and how confident it is—important for high-stakes decisions.

Two types of explanations become feasible: instance-level and global insights: The work enables both local explanations (why did the model make this specific prediction?) and global explanations (which features matter most overall?), providing complementary perspectives for auditing and understanding model behavior.

Efficient exact explanations reduce the need for approximations in responsible AI audits: By making exact Shapley computation practical for this model class, organizations can verify feature importance rigorously rather than relying on faster but less reliable approximation methods, supporting more trustworthy AI deployment.