SHAP values (SHapley Additive exPlanations)

SHAP values explain an individual prediction by attributing it to the contribution of each input feature, answering "how much did this particular feature push this particular prediction up or down, relative to the model's average prediction." They are based on Shapley values, a concept from cooperative game theory originally used to fairly divide a payoff among players based on their contribution, applied here by treating each feature as a "player" contributing to the prediction.

Computing an exact Shapley value requires testing a feature's effect across every possible combination of the other features being present or absent, which is why practical SHAP implementations use approximations: some specialized to be fast for tree-based models like random forest and gradient boosting, others generic enough to work on any model, including a neural network. A defining property is that the feature contributions for a single prediction add up exactly to the difference between that prediction and the model's average output.

SHAP values are a leading technique in Explainable AI, used both to explain individual predictions to end users or regulators and to audit which features drive a model overall by aggregating values across many predictions. They can be computationally expensive on large models, and a large SHAP value shows association with a prediction, not proof that the underlying feature causes the outcome.