| Abstract |
Data streams have become increasingly important due to the massive volume of data generated by recent advances in electronic devices and sensors widely used in smart cities and Internet of Things applications. However, mining continuous data in non-stationary environments with concept drifts poses significant challenges, such as the need for model updates with recent patterns. A common approach for detecting concept drift is based on model performance monitoring, where a drop in performance indicates the occurrence of drift requiring a model update. The Adaptive Windowing (ADWIN) is the most representative method that follows the performance monitoring approach. However, ADWIN and similar methods often rely on labeled data, which may not be readily available in streaming scenarios. We propose ADWIN-U (Adaptive Windowing for Unsupervised Drift Detection), an unsupervised concept drift detector based on the state-of-the-art ADWIN. Our proposal is independent of labeled data for drift monitoring, making it suitable for real-world stream problems where complete labeled data is costly or impractical. Our experimental evaluation demonstrates that ADWIN-U outperforms its supervised version in various domains. Furthermore, we address the limitations of existing evaluation measures for drift detection, which tend to favor approaches that require extensive labeled data. We propose a novel evaluation metric—Balanced Accuracy by the Amount of Requested Labeled Data (BAR). BAR considers the trade-off between accuracy and the proportion of labeled data requested for model updates, favoring accurate detectors with low false alarm rates while minimizing the reliance on labeled data. © The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2025. |
