Shap: Shap value - train/test set

[mlambelho]

First of all,congrats for the amazing shap package @slundberg.

I understand that the following code produces the shap values for every feature in every observation of my model:
explainer = shap.TreeExplainer(my_model)
shap_values = explainer.shap_values(X_train)

Then, it is possible to plot for a single observation the shaps values for every feature:

shap.force_plot(explainer.expected_value, shap_values[0,:], X_train.iloc[0,:], link = 'logit')

However, my question is: how can I evaluate the shaps values of a possible prediction on the test set?
Ideally, I would like to also observe what is the 'importance' of each feature in every single prediction - in other words: which feature has the highest shap value in every single prediction?

[ddearauj]

I believe that if you use your test set you get the shap value for each prediction in your dataset

[mlambelho]

@ddearauj , you mean use the test set when discovering the shap values? Something like this?

explainer = shap.TreeExplainer(my_model)
shap_values_test = explainer.shap_values(X_test)
shap.force_plot(explainer.expected_value, shap_values_test[0,:], X_test.iloc[0,:], link = 'logit')

[slundberg]

That's right. my_model contains data about the training set that it was built on, so TreeExplainer uses that for computing conditional expectations. You pass to explainer.shap_values whatever samples you want to explain, which is often a test set.

[mlambelho]

Thank you! Could you also clarify the correct explanation of what is the base_value? I have read here that it should correspond to the mean of the predictions of your classifier, however, I am using a LightGBM model and unfortunately that is not the case.

[slundberg]

For TreeExplainer it is the mean of the output of the trees, which for logistic regression in LightGBM is the mean of the log-odds predicted by the model. This is different than the mean of the probabilities.

[mlambelho]

Thank you @slundberg ! Completely understood!