ADES: A New Ensemble Diversity-Based Approach for Handling Concept Drift

Abstract

Beyond applying machine learning predictive models to static tasks, a significant corpus of research exists that applies machine learning predictive models to streaming environments that incur concept drift. With the prevalence of streaming real-world applications that are associated with changes in the underlying data distribution, the need for applications that are capable of adapting to evolving and time-varying dynamic environments can be hardly overstated. Dynamic environments are nonstationary and change with time and the target variables to be predicted by the learning algorithm and often evolve with time, a phenomenon known as concept drift. Most work in handling concept drift focuses on updating the prediction model so that it can recover from concept drift while little effort has been dedicated to the formulation of a learning system that is capable of learning different types of drifting concepts at any time with minimum overheads. This work proposes a novel and evolving data stream classifier called Adaptive Diversified Ensemble Selection Classifier (ADES) that significantly optimizes adaptation to different types of concept drifts at any time and improves convergence to new concepts by exploiting different amounts of ensemble diversity. The ADES algorithm generates diverse base classifiers, thereby optimizing the margin distribution to exploit ensemble diversity to formulate an ensemble classifier that generalizes well to unseen instances and provides fast recovery from different types of concept drift. Empirical experiments conducted on both artificial and real-world data streams demonstrate that ADES can adapt to different types of drifts at any given time. The prediction performance of ADES is compared to three other ensemble classifiers designed to handle concept drift using both artificial and real-world data streams. The comparative evaluation performed demonstrated the ability of ADES to handle different types of concept drifts. The experimental results, including statistical test results, indicate comparable performances with other algorithms designed to handle concept drift and prove their significance and effectiveness.