Weighted Conditional Mutual Information Based Boosting for Classification of Imbalanced Datasets
Ákos Utasi · SZTAKI Publication Repository (Hungarian Academy of Sciences) · 2012
This paper addresses the problem of binary classifier learning when the training data is imbalanced, i.e. the samples of the two classes have significantly different cardinality. We investigate two different cost-sensitive approaches in the conditional mutual information (CMI) based weak classifier selection procedure using histogram descriptors. The first method uses CMI for classifier selection, and cost factors are utilized in the construction of the final boosted classifier using support vector machine learning. In the second approach these costs are incorporated into the classifier selection step by weighting the CMI (wCMI). We evaluate the proposed methods in object recognition and detection tasks using two popular histogram-like descriptors. Extensive experiments showed that the proposed methods provide efficient tools to address both problems.