Feature attribution assigns portions of a model output to individual input features. SHAP performs this locally by estimating each feature’s contribution to one prediction, using Shapley values from cooperative game theory. section 9.4.2
Splitting a team prize
Imagine four data scientists sharing a $20,000 competition prize. They evaluate every individual and group coalition, then distribute the total according to each member’s average contribution across coalitions. SHAP applies the same intuition to features and predictions. section 9.4.2
A local SHAP explanation proceeds by:
- Comparing predictions with and without features across possible coalitions. section 9.4.2
