Multi-class classification with 100 classes

User 2568 | 5/6/2016, 12:30:16 AM

I'm working on the Expedia Kaggle competition. This is a multi-class classification problem with 100 classes. I've tried using classifier.create but the results are poor.

  1. How well do classifiers work with such large number of classes.
  2. Is there anything particular I should do
  3. The classes are nominal. Currently they are ints, should I make them strs to indicate they are nominal not ordinal? That is, does this matter?

Comments

User 1207 | 5/16/2016, 6:32:17 PM

@Kevin_McIsaac,

It's likely that with that quantity of classes, you'll need a bit more tuning beyond the default classifier.create. Our algorithms do work well for these models, but it's very likely you'll need good values for your regularization parameters, which the default classifier.create does not necessarily find.

You should be fine if your target column is numeric.

-- Hoyt