Article
Identifying Concept Drift with a Classifier Ensemble Method
The concept drift or the change in the data distribution, hinders the sustainability of the accuracy of machine learning models greatly in dynamic conditions. In this work, an ensemble-based method of successful detection and adaptation of drifts of ideas is proposed. The proposed approach employs a range of classifiers to monitor decreases in the performance and relies on the trend in disagreement or accuracy to detect drift. The ensemble is adaptive as it re-trains or replaces affected models in a case of drift to ensure continued reliability. Both the artificial and real-world data on both test sets indicate that, unlike the traditional single model solutions, our approach delivers a high level of robust classifications and increased sensitivity of the detection. This methodology is well suited to applications in streaming data, where the drift needs to be detected very quickly and accurately to be used in a decision-making system.
Full Text Attachment
