I have a training data for training a classification model. However, i know that this data contains columns (features) that will not be available at test time when I want use the trained model to predict. I do not see any value add from using such features at training time. Note that this is different from a situation where some features may or may not be available at test time, in which case I can see some value of using those features for training a model and then doing some sort of imputation if they were missing at test time. I am not a researcher so wanted to get others thoughts or pointers on this.
1 Answer
$\begingroup$
$\endgroup$
No, there is no value to do so. If you train a logistic regression (for example), for each feature a weight will be associated. If at test time, you do not have one feature anymore, what will you do ? You will put a "fake" feature equal to zero instead ? But the problem is that you taught your model to predict with this feature : this feature has a "signification" for the model. So the performance will be poorer.
