[Interpret model] SHAP values to data frame and custom top values

[victor-mariano-leite]

Is your feature request related to a problem? Please describe.

As far as I've seen, we can only plot the SHAP values when model is trained, it would be useful if we had an parameter to output to a dataframe, instead of a HTML visualization.

And set the ranking threshold of the visualization, that it's hardcoded as 25.

Describe the solution you'd like

output = 'html' (default) parameter, and output = 'df' to output the SHAP mean values as a dataframe, it would simplify the versioning and comparing the values and rankings with other model versions.

Implemented simply like this in issue #970 from SHAP.

top parameter, defaults to 25, on top of SHAP max_display parameter.

[ngupta23]

You can extract the shapley values manually for now. Refer to this notebook (it is towards the end)

shap_values = shap.TreeExplainer(best_model).shap_values(X_train_transformed)

[victor-mariano-leite]

@ngupta23 it's true! Nice, I'll close the issue then