driftwhen

Driftwhen is a term used in data stream analysis to refer to the estimated time at which concept drift begins to affect a data-generating process in a detectable way. It is not a standardized metric and definitions vary across sources, but a common interpretation treats driftwhen as the moment when changes in the data distribution or the relationship between inputs and outputs become statistically significant enough to warrant attention.

Definition and measurement

In practice, driftwhen is derived from a drift score time series produced by a drift detector applied to the stream. The driftwhen time is defined as the first time index at which the drift score exceeds a predefined threshold for a sustained period, such as a minimum window length. This ensures that transient fluctuations do not falsely indicate drift. The exact threshold and window length are typically chosen based on the desired false-positive rate and the characteristics of the application.

Relation to drift detection

Driftwhen relates to established drift detection methods such as DDM (Drift Detection Method), EDDM (Early Drift Detection Method), and ADWIN. These detectors provide scores or statistics that quantify changes in distributions or error rates. Driftwhen uses these signals to pinpoint the onset of drift, rather than simply signaling that drift has occurred, enabling time-aware responses.

Applications and limitations

Driftwhen is useful for scheduling model retraining, triggering alerts, and aligning maintenance activities with the dynamics of a data stream. Its usefulness depends on detector parameters, data noise, and window choices; inappropriate thresholds can yield premature or delayed indications of drift. As with all drift timing metrics, driftwhen should be interpreted alongside other indicators and domain knowledge.

See also: concept drift, drift detection, data streams.