SHAP summary is not supported

[sarveshb-1999]

Cannot calculate SHAP summary for the YDF model using the SHAP library.
Works with the KernalExplainer workaround with a model wrapper, but is very very slow.
Can SHAP support be added to the library?

[rlcauvin]

By "SHAP support", do you mean that you want to use shap.TreeExplainer to explain models created with YDF?

[achoum]

Hi Sarves,

SHAP is a highly requested feature in YDF, and we’re currently developing it 😃.
It will be available in a future YDF version (maybe the next one) for both model and prediction interpretation.

I'll update this issue when it's available.

Details:

Kernel SHAP is a model-agnostic method (i.e., it works for any model; which is great) implemented in the KernalExplainer. However, as you mentioned, it's slow (not so great).
The TreeExplainer mentioned by Roger in the SHAP library uses the Tree SHAP algorithm, which is drastically faster. However, TreeExplainer is not compatible with YDF.
The version of SHAP in development in YDF uses the TreeSHAP algorithm and also supports YDF particular features (e.g., oblique splits and categorical-sets conditions).