A Bayesian active learning framework for a two-class classification problem

TL;DR: The proposed active learning procedure uses a criterion based on the entropy of the posterior distribution of the adaptive parameters to select the sample to be included in the training set to tackle the problem of kernel-based data classification.

Abstract: In this paper we present an active learning procedure for the two-class supervised classification problem. The utilized methodology exploits the Bayesian modeling and inference paradigm to tackle the problem of kernel-based data classification. This Bayesian methodology is appropriate for both finite and infinite dimensional feature spaces. Parameters are estimated, using the kernel trick, following the evidence Bayesian approach from the marginal distribution of the observations. The proposed active learning procedure uses a criterion based on the entropy of the posterior distribution of the adaptive parameters to select the sample to be included in the training set. A synthetic dataset as well as a real remote sensing classification problem are used to validate the followed approach.