Pre CDACI: Prequential Learning Based Concept Drift Detection and Adaptation for Classification of Imbalanced Data Streams

Abstract

Despite classification being one of the most extensively researched problems in data mining and machine learning, the performance of existing models often degrades when trying to deal with data streams due to presence of concept drifts. The problems further increase when the data is imbalanced wherein many instances of minority classes are being overlooked by the classifiers. While evaluating the classifier accuracy in data streams, use of prequential learning provides a true metric of online evaluation. This paper describes a framework that uses an ensemble of classifiers to classify a data stream where each of the classifiers is created using prequential learning technique. The drift is detected comparing the cumulative sum of the data instances to predefined threshold. In order to adapt to the current concept the classifier is retrained on the misclassified instances. The results prove that the framework outperforms some of the well-known classifiers while classifying imbalanced data streams with concept drift.