Improved Ensemble Methods to Solve Multi-class Imbalance Problem Using Adaptive Weights

Nowadays, usage of the internet-of-things and oftenest of mobile phones, the technology, and networks of sensors has led to a huge and sudden increase in the amount of data commonly available in a streaming fashion. Online learning plays a vital role in classification which paves way for adaptive learning algorithm. Online multi-class imbalanced learning is a virgin problem that coalesces to the disputes of both online learning and multi-class imbalance learning. Sometimes the data streams for many applications like disease diagnosis, fraud detection, etc., may result in skewed class distribution. Many works in the literature are focused on online class imbalance learning with two-class problem. In this paper, a novel approach of EMOOB and EMUOB algorithms was proposed that overcome multi-class imbalance problem, where data comes in an online fashion. A comprehensive analysis was made in terms of class imbalance status, data distributions, imbalance rate, and in the performance of classifier. Based on the comprehension gained, two new ensemble methods EMOOB and EMUOB with adaptive weights are proposed in the name of WEOB.