1
$\begingroup$

How do I supposed to deal with the situation when one of the classes is very general, it actually covers a bunch of classes, but other 3 classes I have are very precise. I have a data that has a lot of classes, I just need 3 of them, and I use 4th class as a class for everything else. What problems such method of labeling can cause and how do I deal with those problems? Thank you!

My data has about 170 features and I have about 1500 samples. I plan to use Random Forest classifier.

$\endgroup$

1 Answer 1

1
$\begingroup$

With 170 features and only 1500 examples, I'd be worried about overfitting. So make sure you use a validation set and use some kind of regularization if performance on the training and validation sets differs much.
But regarding the classes, you seem to have 4 classes and care about the distinction between each pair of classes, so using classification with 4 classes seems like the right choice.
One thing you could consider is if that 4th class actually consists of 2 or 3 clearly distinct irrelevant classes, you could classify those. So if your data consists of 5 classes and you threw 2 together because you don't care about those 2, you could try classifying the 5 classes. Giving the algorithm 5 more distinct classes might work better than 4 classes of which 1 is more complicated because it is the combination of 2 classes. However this is highly problem domain specific. If you don't have clearly defined subclasses for that 4th irrelevant class, splitting it isn't going to help you.

$\endgroup$
1
  • $\begingroup$ Thank you! I set a maximum depth of a tree to 8 to avoid over fitting, and it works good. $\endgroup$ Commented Nov 10, 2016 at 9:36

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.