Clustering-Based Optimised Probabilistic Active Learning (COPAL)

Abstract

AbstractFacing ever increasing volumes of data but limited human annotation capacities, active learning approaches that allocate these capacities to the labelling of the most valuable instances gain in importance. A particular challenge is the active learning of arbitrary, user-specified adaptive classifiers in evolving datastreams.We address this challenge by proposing a novel clustering-based optimised probabilistic active learning (COPAL) approach for evolving datastreams. It combines established clustering techniques, inspired by semi-supervised learning, which are used to capture the structure of the unlabelled data, with the recently introduced probabilistic active learning approach, which is used for the selection among clusters. The labels actively selected by COPAL are then available for training an arbitrary adaptive stream classifier. The performance of our algorithm is evaluated on several synthetic and real-world datasets. The results show that it achieves a better accuracy for the same budget than other recently proposed active learning approaches for such evolving datastreams.KeywordsProbabilistic active learningSelective samplingEvolving datastreamsNonstationary environmentsConcept driftAdaptive classificationClustering