Network Digital Twins with Concept Drift Detectors: A Comparative Analysis

In a Network Digital Twin (NDT), unsupervised concept drift detection triggers crucial model updates. This paper evaluates Kolmogorov-Smirnov Windowing (KSWIN), Adaptive Windowing (ADWIN), and Page-Hinkley (PH) on a simulated traffic stream. Our sensitivity analysis reveals that window- based detectors are highly susceptible to the alignment between buffer sizes and the stream's stationary periods. KSWIN and PH showed apparent optimal performance at a 7 100 window merely by synchronizing with the distribution's duration, but failed completely at 9 000. ADWIN remained overly conservative, risking under-detection. Calibrating unsupervised detectors requires prior knowledge of the network's stationarity to prevent critical failures in operational NDTs.