Concept Drift

Concept drift studies how changing data patterns or target relationships affect models after deployment, especially in streams such as network traffic, malware, and time-series forecasts. Methods detect changes using prediction errors, statistical tests such as MMD and CUSUM, or shifts in representations, then respond through retraining, online updates, ensembles, replay, or memory of recurring regimes. Evaluation uses synthetic streams with controlled abrupt, gradual, and recurring shifts alongside chronological real-world benchmarks, often measuring detection delay, accuracy, labeling needs, and computational cost. A central challenge is balancing quick adaptation against retaining useful knowledge: model updates can hide drift from error-based detectors, while weak signals, delayed labels, false alarms, and limited budgets complicate reliable intervention.