Active learning for online nonlinear Neyman-Pearson classification [Çevrimiçi doǧrusal olmayan Neyman-Pearson sınıflandırması için aktif öǧrenme]

Official URL: https://dx.doi.org/10.1109/SIU55565.2022.9864669

Neyman-Pearson (NP) classification framework is suitable for solving binary classification problems with asymmetric error costs such as network intrusion detection and medical diagnosis. In these kind of applications, type I (detecting non-target as target, false positive) and type II (detecting target as non-target, false negative) errors have different consequences. In this paper, we propose an active learning method for online context tree based ensemble NP classifiers. Proposed method prioritizes training samples that have high uncertainty (greater than a constant threshold) among different classifiers of the ensemble model. We report the performance of the proposed active learning method by measuring the moving true positive rates (TPR) and NP scores with respect to the number of samples used in learning. Experiments are carried out on 4 different datasets and proposed model was compared with random sampling method, where new samples are selected randomly from the training set. In addition, we also show that in order to satisfy target false alarm rate of the NP problem, we need to sample training set with and exploration probability, independent from uncertainty measurement.