0
$\begingroup$

I have a dataset I believe is easily clustered into a few groups, however, a bunch of junk features interfere with clustering. Is there a method of eliminating these bad features within this dataset?

For example, if you consider hot sauce and want to group different brands into hot, medium, and mild, the color and acidity of the hot sauce will probably tell you something important, but how quickly a hot sauce evaporates may be "junk data" which adds white noise to your graph of hot sauce samples. I'm looking for a way to get rid of data that makes it harder to group these hot sauces into the categories you have already determined ( such as hot, medium or mild) for each sample.

$\endgroup$
  • $\begingroup$ Add a visualization to your question. Maybe the data doesn't cluster as easily as you believe. $\endgroup$ – Has QUIT--Anony-Mousse Oct 19 '15 at 22:15
  • 1
    $\begingroup$ @JamesBeezho I've edited your question a bit to use more standard terminology. If I misinterpreted it, please revert it and let me know. $\endgroup$ – Dougal Oct 20 '15 at 9:33
  • 1
    $\begingroup$ Thanks for the update and thanks @Dougal for the edit. What I think should be clarified before I can vote to reopen is what exactly is "predefined groups" in the title and "categories you have already determined" in the last sentence. Do you mean that you want to select a subset of features that yields good unsupervised clustering (option 1) or that you want to select a subset of features that allows good supervised decoding of the group membership (option 2)? In your example, do you assume that you know the group each sauce belongs to (this will be option 2), or that you don't know it? $\endgroup$ – amoeba says Reinstate Monica Oct 20 '15 at 11:09
  • $\begingroup$ The question was in regard to option 2. Thank you Dougal for the correct terminology. Feature selection is what I was looking for. $\endgroup$ – James Beezho Oct 21 '15 at 6:59
0
$\begingroup$

Do you already know the hotness of the hot sauces? If so, you want to do discriminant analysis NOT cluster analysis. If you are doing discriminant analysis, most techniques have methods for testing individual explanatory variables. For instance, if you are doing a linear discriminant function, MANOVA can be used initially to eliminate non-significant variables. Partition trees, Neural nets, kernel discriminant analysis, etc. often have built in (often randomization-based) tests to see if individual variables are junk and can be removed.

$\endgroup$

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.