AssertionError: shap interaction values of binary classifiers must add up to 0.0 for each observation and feature pair

[mgelsm]

Describe the bug
When inspecting a binary classifier, the raw_shap_tensor of class 0 does not equal to -raw_shap_tensor of class 1.
It appears that the absolute difference can reach up to 10^-2.

Bug rises in function raw_shap_to_df

To Reproduce
Steps to reproduce the behavior:

1. Go to the NHO facet modelisation and run the 4-Facet-modeling-NewAPI
2. Try fitting the LearnerInspector instance
3. See error:

Expected behavior
A clear and concise description of what you expected to happen.

Screenshots

Desktop (please complete the following information):

• OS: [e.g. iOS] IOS
• Browser: Brave

[j-ittner]

Thanks for spotting this!

By definition, both SHAP tensors obtained for a binary classifier should add up to 0.0 for each observation and feature.

Your example provides evidence that totals may deviate by as much as as 0.01 due to imprecisions in the SHAP explainer's approach for estimating SHAP values.

This is now addressed by PR #24.

The fix is not to raise an exception if the totals are not 0.0, but to log a warning instead, stating the range of observed totals. As long as these totals are small (e.g., less than 0.05, corresponding to 5%pt probability), it should be safe to ignore these warnings.