Semi supervised adaptive framework for classifying evolving data stream

Abstract

Most of the approaches for classifying evolving data stream divide the stream into fixed size chunks to address infinite length and concept drift problems. These approaches suffer from trade-off between performance and sensitivity. To address this problem, existing adaptive sliding window techniques determine chunk boundaries dynamically by detecting changes in classifier error rate which requires true labels for all of the data instances. However, true labels are scarce and often delayed in reality. In this paper, we propose an approach which determines dynamic chunk boundaries by detecting significant changes in classifier confidence scores using only limited number of labeled data instances. Moreover, we integrate suitable classification technique with it to propose a complete semi supervised framework which uses dynamic chunk boundaries to address concept drift and concept evolution efficiently. Results from the experiments using benchmark data sets show the effectiveness of our proposed framework in terms of handling both concept drift and concept evolution.