CDACI: Concept Drift Detection and Adaptation to Classify Imbalanced Data streams

Abstract

Non-stationary data streams usually are affected by the phenomenon of concept drift. A drift undetected leads to the drastic drop in the classifier’s accuracy. Detecting this drift and adapting to it becomes a challenging task while classifying data streams. The problem further increases in imbalanced data streams. This paper proposes a model CDACI: Concept Drift Detection and Adaptation to Classify Imbalanced Data streams to classify imbalanced data streams in the presence of concept drift. CDACI uses a cosine similarity function to detect concept drift between subsequent windows. The classifier learns the new concept and forgets the earlier concepts when the drift is detected otherwise the classifier continues to train on the incoming data while remembering previous concept, this behaviour of the classifier is determined by the optimized threshold value. The results show that CDACI is successfully able to adapt to new concepts and classify imbalanced data streams.