SHAP

Shapley values originate in cooperative game theory, where contributions are allocated by considering a participant's marginal effect across possible coalitions. SHAP applies this idea to model explanations, treating features as participants and defining a value function for subsets of available features. Attributions sum to the difference from the chosen baseline under the method's assumptions. Different SHAP algorithms exploit model structure or approximate the computation, and their handling of feature dependence can differ. The resulting contribution describes the explanation's defined prediction game; it is not automatically a causal effect or an intrinsic property of a feature independent of the reference population.

The practitioner chooses an algorithm appropriate to the model and an explicit background dataset. They inspect local attributions and aggregate them carefully for broader analysis, preserving the distinction between signed effects and absolute magnitude. Useful artifacts include the baseline definition, explanation settings and checks on representative cases. Correlated features require attention because credit can shift depending on assumptions about feature absence. The team verifies whether explanations answer the intended question and whether changes in background data alter conclusions that stakeholders might otherwise treat as stable facts.

A precise-looking attribution can hide strong assumptions about the baseline and correlated inputs. Some computation methods are approximate, and aggregate importance can obscure subgroup behavior or interactions. SHAP does not establish fairness, causal validity or factual accuracy. Quality checks should state the explanation variant and reference data, compare plausible alternatives and avoid turning a model association into a recommendation for real-world intervention without further evidence.