Inproceedings,
Cost-sensitive Multiclass Classification Risk Bounds
ICML, page 1391--1399. (June 2013)
Abstract
In this paper we present 0-1-like-risk bounds for multiclass cost-sensitive classifiers minimizing losses from an often-used family surrogate losses. To this end, we calculate an analytic expression that describe how the 0-1-like-risk-convergence rate of a classifier depends on the surrogate loss chosen while making less assumptions on the surrogate losses than previous work. We also show that calculating the values of the resulting expression is as easy as calculating these values for binary classification, and derive that some well-known losses share the same rate of convergence as their ``truncated versions'', a fact previously observed only through examples and for some of these losses.
