Loss functions to combine learning and decision in multiclass problems

Abstract

The design of structures and algorithms for non-MAP multiclass decision problems is discussed in this paper. We propose a parametric family of loss functions that provide the most accurate estimates for the posterior class probabilities near the decision regions. Moreover, we discuss learning algorithms based on the stochastic gradient minimization of these loss functions. We show that these algorithms behave like sample selectors: samples near the decision regions are the most relevant during learning. Experimental results on some real datasets are also provided to show the effectiveness of this approach versus the classical cross entropy (based on a global posterior probability estimation).