Computing Conditional Shapley Values Using Tabular Foundation Models

Shapley values are computationally expensive but important for AI transparency: Shapley values help explain AI decisions by showing how much each data feature contributes to a prediction, but calculating them traditionally requires enormous computational effort, especially when features are related to each other.

A new approach using foundation models dramatically speeds up explanations: The research demonstrates that tabular foundation models (specifically TabPFN) can compute Shapley values much faster than existing methods by avoiding the need to retrain models repeatedly—instead using "in-context learning" to estimate values without retraining.

The new method performs better or comparably to existing approaches at a fraction of the cost: Testing on both simulated and real datasets showed that TabPFN either outperformed state-of-the-art methods or came very close while requiring significantly less computational time, making AI explanations more practical for real-world use.

This advancement makes explainable AI more accessible: By reducing the computational burden of generating Shapley value explanations, this work helps make it feasible for organizations to explain their AI systems' decisions more routinely, supporting responsible AI practices and regulatory compliance.

Further optimization is possible: The authors note that tabular foundation models could be specifically refined for this task, suggesting even better performance and efficiency gains may be achievable in the future.