1
$\begingroup$

I learned that building an ensemble model using stacking is done by training a meta-model on the predictions of $n$ other models in order to combine the predictions and try to enhance the performance. But can we in addition add the original input features to teach the model a relation between specific input features and the models predictions, hence, increase the accuracy?

$\endgroup$

1 Answer 1

1
$\begingroup$

You certainly can. Whether you train your meta-model on just the base model predictions, predictions + original data, transformations of one or both of these, etc. can be viewed as just another hyper-parameter you can tune in your meta-model.

In Python, StackingClassifier has an argument to allow you to do exactly what you've suggested (passthrough = True).

When you view your base model out-of-sample predictions as "just another feature", the possibilities are pretty endless.

$\endgroup$

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.