Clever Geek Handbook

Large number of classes

I am working on a multi-class model with a huge number of classes (about 3500). Could a large number of classes affect the performance of my model? I would like to use SVM and Random Forest. Does anyone know if there is any class limit for these methods? thanks in advance

+3


source to share


1 answer


Yes, this can lead to performance gains, especially since most libraries solve the multiclass problem using a combination of binary problems. There are different strategies (one-vs-all, one-vs-one, winner-take-all, etc.), and you should try and see which ones are effective enough for you (assuming you have control over the strategy).

https://en.wikipedia.org/wiki/Support_vector_machine#Multiclass_SVM



A dirty hack that has sometimes worked for me in the past is to treat the problem as a regression problem, not a multiclass problem, but that might not be valid in your case, I should have seen the problem say in detail.

+2


source






More articles:


All Articles