Correntropy in Data Classification

Abstract

AbstractIn this chapter, the usability of the correntropy-based similarity measure in the paradigm of statistical data classification is addressed. The basic theme of the chapter is to compare the performance of the correntropic loss function with the conventional quadratic loss function. Moreover, the issues related to the non-convexity of the correntropic loss function are considered while proposing new classification methods. The proposed methods incorporate the correntropic loss function via the notions of convolution smoothing and simulated annealing optimization algorithms. Two nonparametric classification methods based on the correntropic loss function are proposed and compared with the conventional parametric and nonparametric methods. Specifically, the classification performance of the proposed artificial neural network-based methods are not only compared with their conventional counterparts but also with the kernel-based soft margin support vector machines. Experimental studies with Monte Carlo-based simulations show the validity of the proposed methods in the data classification.KeywordsStatisticalclassificationCorrentropyConvolution smoothingSimulated annealingArtificial neural networksSupport vector machines