Comparison based analysis of window approach for concept drift detection and adaptation

Abstract

In the non-stationary data stream distribution, concept drift occurs due to change in patterns with respect to time. It is necessary to identify drift in the data stream during the early stage. One way to explore the change in patterns is windowing, where two windows compare to find the difference in data distribution. In the two-window-based methods, the concept drift may occur much before the incoming window. The current window will wait to compare with a new incoming window’s data distribution for drift detection. It may lead to delay in detection, increasing misclassification error, and decreasing classification accuracy. The paper proposes DD-SCC-I and DD-KRC-I, incrementally adaptive single-window-based drift detection methods, to overcome the above issue. These methods localize the concept change by finding the correlation between attribute vectors. The proposed work deals with multi-dimensional data, binary-class classification, and multi-class classification problems. An improved two-window-based concept drift detection methods, DD-SCC-II and DD-KRC-II, are built to find drift using the same correlation. Further, the comparison is made among proposed methods in terms of the number of drift detected and drift detection times to demonstrate the behavior of methods. These proposed methods compare with state-of-the-art methods using real-time and synthetic data sets. The evaluation result shows DD-SCC-I and DD-KRC-I detect early drift with an increase in average rank of 4.18 and 4.56, respectively.