Classifier Learning for Imbalanced Data

This work discusses the theoretical abilities of three commonly used classifier learning methods and optimization techniques to cope with characteristics of real-world classification problems, more specifically varying misclassification costs, imbalanced data sets and varying degrees of hardness of class boundaries. From these discussions a universally applicable optimization framework is derived that successfully corrects the error-based inductive bias of classifier learning methods on image data within the domain of medical diagnosis. The framework was designed considering several points for improvement of common optimization techniques, such as the modification of the optimization procedure for inducer-specific parameters, the modification of input data by an arcing algorithm, and the combination of classifiers according to locally-adaptive, cost-sensitive voting schemes. The framework is designed to make the learning process cost-sensitive and to enforce more balanced misclassification costs between classes. Results on the evaluated domain are promising, while further improvements can be expected after some modifications to the framework.