Multiclass#

OVA - one versus all#

If have k classes, train k classifiers, each of which predicts yes/no is the data point this class

AVA - all versus all#

If have k classes, train k choose 2 classifiers, each of which says is it more one class or the other. Select class with most votes. Could also do binary tree.

Excercises#

  • Derive OVA, AVA error bounds