Missingness

Also known as local accuracy, this axiom guarantees that the sum of all feature attributions exactly equals the difference between the model's prediction and the baseline value. When combined with missingness, it ensures a complete and faithful decomposition of the output. No contribution is lost or double-counted.

If a model changes so that a feature's marginal contribution increases or stays the same across all subsets, that feature's SHAP value must not decrease. This property ensures that features which become more important to the model are assigned higher attribution, maintaining logical stability across model retrains.

The expected model output computed over the background dataset, representing the prediction when no feature information is known. Missingness requires that features absent from the input receive zero attribution, so the baseline serves as the starting point from which present features deviate.

This property guarantees that the explanation model matches the original model's output exactly for the specific instance being explained. It works in tandem with missingness: features present in the input contribute their SHAP values, while absent features contribute nothing, summing precisely to the prediction.

The class of explanation models that express a prediction as a linear sum of individual feature contributions relative to a baseline. Missingness is a required property of all additive feature attribution methods, ensuring that:

• Features not in the input receive zero attribution
• The sum of attributions reconstructs the prediction

The mathematical foundation for Shapley values, studying how groups of players form coalitions and distribute payoffs. Missingness maps directly to the game-theoretic principle that players not participating in a coalition receive zero payout, ensuring fair distribution only among active participants.