The static thresholds derived from the primary validation timeframe (window) are a common method for the detection of reconstruction-based concept drift. In prolonged periods of data streams, progressive changes in reconstruction accuracy frequently lead to misalignment, giving rise to repeated false alarms and inconsistency in detection behavior. This study introduces a modular lightweight adaptive thresholding strategy for the Autoencoder-Based Drift Detection Method (AEDDM) by integrating an Exponentially Weighted Moving Average (EWMA) mechanism into the batch-level decision process, without modifying the original model architecture. Rather than constituting a standalone framework, the proposed method functions as a modular enhancement to the decision layer of the AEDDM pipeline. The proposed solution is validated using a synthetic Gaussian stream together with the ELEC2 and NSL-KDD datasets. The finding demonstrates that the EWMA-based approach effectively eliminates false alarms without compromising responsiveness under abrupt changes, achieving zero-latency on NSL-KDD compared to static thresholds that produced 22 warnings and 25 false alarms in a stationary stream. Findings from this study suggest that adaptive thresholding alone significantly leads to the enhancement of detection performance in reconstruction-driven drift on a real-time stream.
Rum et al. (Thu,) studied this question.