Adaptive online learning for classification under concept drift

Abstract

In machine learning and predictive analytics, the underlying data distributions tend to change with the course of time known as concept drift. Accurate labelling in case of supervised learning algorithms is essential to build consistent ensemble models. However, several real-world applications suffer from drifting data concepts which leads to deterioration in the performance of prediction systems. To tackle these challenges, we study various concept drift handling approaches which identify major types of drift patterns such as abrupt, gradual, and recurring in drifting data streams. This study also highlights the need for adaptive algorithms and demonstrates comparison of various state-of-the-art drift handling techniques by analysing their classification accuracy on artificially generated drifting data streams and real datasets.