Learning of concept drift and multi class imbalanced dataset using resampling ensemble methods

Abstract

In modern days, very often usage of mobile phones paves way for advanced technologies which includes Internet-of-Things (IoT), wearable technology and big data. As the technology grows, huge volume of data with its complexities also increases rapidly. Flooding of data leads to combat in terms of online class imbalance problem and concept drift. Class imbalance problem is one of the issues in which number of class labels is not balanced and also majority classes are given more importance than the minority class. This type of situations leads to none accurate classification of data. Spam filtering, Fault detection in Engineering industry, Disease diagnosis are few applications where multiclass imbalance with concept drift makes prediction challenging. In this paper, a novel approach of Concept Drift Detector and Resampling Ensemble (CDRE) algorithm was proposed to overcome the problem of concept drift in multi-class. Misclassification occurs sometimes due to imbalance ratio and data distribution. Detailed analysis was done based on different levels of imbalance ratio and data distribution. There is decline in accuracy when multi-class problem suffers from concept drift also. When compared to normal multi-class imbalance problem, class imbalance problem with concept drift is analyzed. Concept Drift Detector and Resampling Ensemble (CDRE) algorithm was implemented to deal multi-class problem with concept drift. CDRE algorithm shows better results in recall, precision, F-measure on an average 85% when compared with algorithm without optimization.