SHAP: Checking Additivity
Intermediate
The problem, from first principles
SHAP values must add up to the difference between the prediction and the baseline prediction. This check is the efficiency property of Shapley values, and it is a quick way to catch a bug in an attribution method.
From theory to code
Implement additivity_gap(phi, f_x, f_base), returning the sum of the attributions minus the output difference.
Constraints
- Returns a Python float.
Hints
Hint 1
Sum the attributions and subtract f_x - f_base.
Keywords
research-papersclassical-mlinterpretabilityshap
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.