Interpreting SHAP summary and dependence plots
SHapley Additive exPlanations (SHAP) is a collection of methods, or explainers, that approximate Shapley values while adhering to its mathematical properties, for the most part. The paper calls these values SHAP values, but SHAP will be used interchangeably with Shapley in this book. However, it must be noted that the authors of SHAP took a few liberties with the properties. For instance, some explainers don't comply with the dummy property and leverage reference background data to simulate missing values. Despite these issues, because of SHAP being grounded in other solid properties, it's still better than alternatives studied in Chapter 4, Fundamentals of Feature Importance and Impact.
It has three properties that are loosely based on Shapley's:
- Local accuracy: Equivalent to Shapley's efficiency property.
- Consistency: Encompasses additivity and substitutability axioms, and, in theory, dummy as well...
