1
$\begingroup$

Is there any general approach to approximate how a missing feature will impact the prediction performance of a regression model?

For example, if I train a model using 10 features, but want to make predictions afterwards only providing 8 of those 10 features - simply imputing the missing values -, how does that (statistically) affect my target metric. You could also think of it like some kind of confidence decline of the predictions.

I wondered if there is some approach like deriving / estimating the deterioration of performance from feature importances, or similar. I did not find any resources about that though, nor could I come up with some meaningful idea myself so far.

$\endgroup$

1 Answer 1

1
$\begingroup$

I would just remove each feature individually and see how that effects performance on my test set. No need to get fancy.

$\endgroup$
1
  • $\begingroup$ I wanted to avoid that, as the feature-space in theory could increase and multiple features could be missing. This then quickly leads to hundreds of combinations for which the model needs to be trained. Therefore I was hoping to find some more generalisable approach. $\endgroup$
    – Niggl
    Mar 13, 2020 at 7:42

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.