Precision of Individual Shapley Value Explanations

Shapley value explanations are less reliable for unusual cases: The paper demonstrates that Shapley values—a popular tool for explaining AI predictions—provide less precise explanations for observations that fall outside the typical range of training data. This means explanations for unusual or edge-case predictions should be treated with extra caution.

Precision varies by individual observation, not just on average: While previous research compared Shapley value methods based on average performance across many predictions, this work reveals that precision varies significantly from one prediction to another. Some individual explanations are inherently more trustworthy than others, depending on where the data point falls in the distribution.

Practitioners need awareness of this limitation: The authors emphasize this finding is "crucial knowledge for Shapley values practitioners," suggesting that professionals using these explanations should actively assess whether each observation is typical or unusual before relying on its explanation. This requires additional due diligence beyond simply applying the method.

Statistical theory predicts this pattern: The reduced precision for outliers aligns with established statistical principles, but the paper notes this connection "has not been systematically addressed in the Shapley value literature"—meaning the AI explainability field may have overlooked a known statistical phenomenon that should inform how these tools are used responsibly.