Are SHAP calculations in ExplainerDashboard aggregate? What are the ideal sizes for X_background and X when evaluating neural net models?

[Menamonmon]

I have two questions here:

CanExplainerDashboard be used to examine the influence of features on a single prediction or does it only provide accurate explanations at an aggregate level for the model as a whole? If that's possible, how big should the X and X_background datasets be when passing them to ClassifierExplainer.

How big should the X_background be if I want to ensure accuracy of the SHAP values for a large training dataset (millions of rows for training)?
If SHAP calculations only give valuable insights at an aggregate level, similarly, how big should the X dataset when generating the ClassifierExplainer?

Would really appreciate some examples for clarifications on this? So far, I saw in the examples that use the training data as X_background and the testing data as X but in my case, I have a very large dataset and want to optimize performance vs accuracy and also want to see if getting explanation for individual predictions are possible. Thank you in advance!

[oegedijk]

Hi @Menamonmon, the idea of SHAP values are specific to the example, but can then be aggregated to get global insights into your model.

In terms of the size of X_background, I think this can be pretty small and still get decent results (tens of examples), but this is probably a question better asked on the shap library github.