3
$\begingroup$

Let's assume that I have a classification problem and all my features are categorical data.

I have missing data (and I do not want to do any imputation).

Also, I know that I will have some unseen data (at my test data) at some of my features.

My question is the following:

Should I encode the missing and unseen data (of the test set) into the same class or to different ones?

Which is the most common practice and why?

$\endgroup$
3
  • 2
    $\begingroup$ Could you explain what you mean by "unseen data" and how they differ from missing? $\endgroup$
    – Leevo
    Dec 23, 2019 at 16:10
  • 1
    $\begingroup$ @Leevo, thank you for your question. Let's say that we process bank statements and one feature we have in our dataset based on the OCR data we extracted from the bank statements is about the name of the bank. At the test set, sometimes we may not be able to extract or identify the bank's name so we will have missing data. However, also we may have extracted and identified it but it may be the first time we see this bank name - this bank name did not occur at all in the training data; in this case we have "unseen" data. Does this make sense now? $\endgroup$
    – Outcast
    Dec 23, 2019 at 16:21
  • 1
    $\begingroup$ @Leevo, the question is, should I put these two different data categories in the same class or in different classes when I encode these data at the test set? $\endgroup$
    – Outcast
    Dec 23, 2019 at 16:22

1 Answer 1

0
$\begingroup$

This depends at least a little on your model type and encoding strategy. It sounds like you'll be one-hot encoding, with possible level/dummy for "missing" and/or for "new". Given you say you won't impute, I guess you'll be using tree-based models?

If you have a dedicated level for unseen categories, then the trained model will have no idea what to do with it. Results will vary among errors in model build, zero input of the dummy feature, and random contributions from the dummy feature.

Assigning new categories to the missing category is a little better, but will depend on how you treat missing values in your model training (but you've said you don't want to impute). At least it gives a consistent result. This may even be the best choice, depending on your data.

In your elaboration you mention bank names, which suggests an alternative (maybe several).
Perhaps these unseen banks will have some commonality with banks in the training set (they're small and local maybe?); if so, you could try to coarsen the categories in the train set so that these new ones will fall nicely into those categories (and then encode all future unseen banks into that new category; make a note to reexamine this process when you retrain!).
If your training data has lots of small levels, it can also make sense to lump the rare categories together (to help avoid overfitting); in that case you can just drop the new levels into that one. Again depending on your data this might work well (small and local would also be rare, so this accidentally recovers the last idea in this case).

$\endgroup$

Your Answer

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

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