Sensitivity-Based Optimization of Unsupervised Drift Detection for Categorical Data Streams

Real-world data streams are rarely characterized by stationary data distributions. Instead, the phenomenon commonly termed as concept drift, threatens the performance of estimators conducting inference on such data. Our contribution builds on the unsupervised concept drift detector CDCStream, which is specialized on processing categorical data directly. We propose a cooldown mechanism aiming at reducing its excessive sensitivity in order to curb false-alarm detections. Using practical classification and regression problems, we evaluate the impact of the mechanism on estimation performance and highlight the transferability of our mechanism on other detection methods. Additionally, we provide an intuitive means for tuning the sensitivity of drift detectors. While only marginally improving the unaltered form of the detector on publicly available benchmark data, our mechanism does so consistently in almost all configurations. In contrast, within the context of another real-world scenario, almost none of the tested drift-detection-based approaches could outperform a baseline approach. However, potentially false-alarm detections are reduced drastically in all scenarios. ... mehrWith this resulting in a cutback in signals for refitting estimators, while maintaining a better or at least comparable performance to vanilla CDCStream, compute infrastructure utilization could be economized further.