Model Error Analysis

[lcrmorin]

After training a model, it is possible to get explaination with shapley value by activating the contrib option in the model. I was wondering if there is a natural way to aggregate features contribution for model error analysis.

That is, we are given a model calibrated on X_train, y_train, and some test set X_test, y_test on which we predict probabilities and shapley values. For model error analysis - knowing which feature cause bad predictions - is there a natural way to aggregate shap values, weighted by the difference between predictions and true values ?

Although not really needing some devellopement, It would be a nice usecase to document.

Edit: I reckon this might be a question about Shapley Values more than lgbm; Still, as lgbm now include a very fast shapley algo this might be a relevant usecase.