3
$\begingroup$

I predict a continuous variable by taking the average of $N$ model predictions. The models are different in terms of their functional form, i.e. a tree model, a neural net, etc.

Is the average SHAP value for some input variable across those $N$ models equal to the SHAP value I would get by treating the aggregate model as a single model? Put differently, if I wanted to derive the SHAP value of an input variable for my aggregate (average) model prediction, is taking the average of the individual SHAP values for that respective input variable across models the correct approach?

$\endgroup$

1 Answer 1

2
$\begingroup$

To answer my own question: Yes, apparently linearity is one of the attributes of SHAP values. I found the answer in this github issue thread: enter link description here. The response I am referring to is the one by the "inventor" of SHAP values for model interpretation, so I guess it's trustworthy.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.