Active Learning by Clustering for Drifted Data Stream Classification

Usually, during data stream classifier learning, we assume that labels of all incoming examples are available without any delay and they are used to update employing predictive model. Unfortunately, this assumption about access to all class labels is naive and it requires relatively high budget for labeling. It causes that methods which can train data stream classifiers on the basis of partially labeled data are highly desirable. Among them, active learning [1] seems to be a promising direction, which focuses on selecting only the most valuable learning examples to be labeled and used to produce an accurate predictive model. However, designing such a system we have to ensure that a chosen active learning strategy is able to handle changes in data distribution and quickly adapt to changing data distribution. In this work, we focus on novel active learning strategies that are designed for effective tackling of such changes. We propose a novel active data stream classifier learning method based on query by clustering approach. Experimental evaluation of the proposed methods prove the usefulness of the proposed approach for reducing labeling cost for classifier of drifting data streams.