SHAP Feature Attribution

SHAP feature attribution uses Shapley values from cooperative game theory to allocate a model prediction among input features by comparing predictions across feature subsets. KernelSHAP estimates values for black-box models, while TreeSHAP and DeepSHAP exploit tree and neural-network structure; other variants use sampling, conditional or hierarchical feature groups, and specialized units for images and time series. Researchers evaluate explanations for fidelity to model behavior, stability across samples and retraining, and usefulness in applications such as healthcare, security, and finance. A central tension is that correlated features can make rankings unstable or non-unique, results depend on how missing features are represented, and faithful attribution alone does not establish causal importance.