While working on Kaggle Competition, I ended up with 11,726 columns which are mostly "dummies" (one hot encoding). Is this too many?

I know that we need to find out which features are relevant, but not sure how to do this.

  • $\begingroup$ did you try asking this question in the competition forum? $\endgroup$ – oW_ Feb 20 '19 at 23:43

Your solution will depend on a couple of factors. One is what type of model you are using. If you are using something that automatically calculates feature importances then you could simply look at these (or take a more balanced look using permutation importances).

While you could look at feature importances, with ~11,000 possibilities this is going to be pretty difficult. The main focus should be to cut down these features into something more manageable, do you really need one hot encoding? Without knowing more about the dataset I can't provide much more advice.

| improve this answer | |
  • $\begingroup$ mlbook.explained.ai looks like a great resource. Do you have any others? $\endgroup$ – B Seven Feb 21 '19 at 18:14
  • $\begingroup$ @BSeven what do you want to look at in particular? $\endgroup$ – HFulcher Feb 21 '19 at 19:25
  • $\begingroup$ I'm interested in ML for Time Series. I liked your link because it has great content and is easy to understand. $\endgroup$ – B Seven Feb 21 '19 at 21:19
  • $\begingroup$ I don’t know of any content for Time Series off the top of my head but will take a look for you. $\endgroup$ – HFulcher Feb 21 '19 at 21:23
  • 1
    $\begingroup$ @BSeven this looks a good place to start. If you feel I answered your question adequately please mark it as answered :) $\endgroup$ – HFulcher Feb 22 '19 at 16:41

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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