Dear All,
Does anybody know some works or system methods on the choice of the appropriate Machine Learning classifier, which is not trial and error-based?
At present, I am using various classifiers, but I would like to justify the choice of my classifier with criteria a bit more scientific than just accuracy.
Thank you very much in advanced for any help.
Cheers,
Gioele
The more information you acquire about the data (prior information, data distribution) and the PR problem at hand, the higher the chance to select a reasonable good classifier. Roughly speaking, the No-Free lunch theorem states that there is no difference between a bad and a good classifier, there is at least one target function (the one being learned) for which random guessing is a better algorithm. If you have to choose one classifier, try to get an insight into the data and the problem you are trying to solve. Avoid over-fitting through a regularization process.
Thank you very much for the reply.
@ Imen Thank you for the reply. I am working with a "large" (up to 50'000 points) labelled dataset and do discriminative learning. At the moment I am trying out Decision trees and ensembles, but only because they are the fastest to train.
@Bekir
Thank you for the reply. It is really interesting, and good reference. However, I am more in need of a workflow that helps me on the selection of a classifier more than a classifier itself.
Regards,
Gioele
- Badges
- Science topic