ADWIN

ADWIN, short for Adaptive Windowing, is an online change-detection algorithm for data streams introduced by Albert Bifet and Ricard Gavaldà. It maintains a variable-length window of the most recent data points and monitors the window for statistically significant changes in the mean of the stream. The goal is to detect concept drift—when the underlying data-generating process changes—without requiring labeled data or prior knowledge of when changes occur.

The core idea is to keep a window W whose size adapts over time. After each new element arrives, ADWIN tries to split W into two subwindows, W0 and W1, at every possible boundary and compares their means. Using a bound derived from Hoeffding’s inequality and a user-specified confidence parameter delta, it determines whether there is statistically significant evidence of a change between the two parts. If such evidence is found, ADWIN shrinks the window from the left by removing the oldest elements until the bound is no longer violated. This way, the window continuously adapts to the current concept and discards outdated information.

Key properties include online operation, unsupervised drift detection, and probabilistic guarantees: with probability at least 1 - delta, the algorithm avoids excessive false alarms in stable data. The delta parameter controls the trade-off between sensitivity to drift and robustness to noise. ADWIN can be used as a component in adaptive classifiers and stream learning systems to trigger model updates or retraining when drift is detected.

Applications of ADWIN span real-time analytics, sensor networks, and any scenario requiring automatic, parameter-free drift detection in evolving data streams.

---