Video Upload Options

Shapley Additive Explanations (SHAP) is a unified framework for interpreting the predictions of machine-learning models by assigning each input feature a real-valued importance (a SHAP value) for a specific prediction ^[1]. SHAP values are derived from the Shapley value concept of cooperative game theory, in which the contribution of a player is the average marginal contribution of that player over all possible coalitions of other players ^[2]. In SHAP, each feature is treated as a player and the model prediction as the game payout, so that the explanation of a particular prediction is decomposed into a baseline value plus the sum of the feature-attribution values ^[1]. The framework unifies a family of additive feature-attribution methods—including LIME, DeepLIFT, and tree-interception methods—under a common set of desired properties: local accuracy, missingness, and consistency ^[1]. SHAP distinguishes itself by providing a theoretically unique and consistent attribution under these properties, and by supporting both local explanations of individual predictions and global summaries of feature importance ^[3].