Concept Drift

Definition

Concept drift is a change over time in the relationship between the inputs and the thing being predicted, so a model or rule that once worked gradually stops working. Data drift, a related idea, is a change in the inputs themselves even if the relationship stays the same.

In plain terms

Worlds change: customer tastes, competitors, laws, and the platforms that produce the data. Drift can be sudden (a rule change), gradual (a slow shift in behaviour) or recurring (seasons). It is easy to miss, because the model keeps producing confident answers.

Detection needs monitoring on two fronts: how the inputs are distributed, and, once outcomes are known, how the performance develops. Remedies include retraining on recent data, giving recent data more weight, and testing across different periods before trusting a result. Remember that outcomes often arrive late, which delays the alarm.

Example

A rule flags items as “rising” when they gain more than 100 ratings in 30 days. Its precision, measured afterwards each quarter (synthetic):

The fall begins in Q3, when the platform changed how ratings are displayed and counted. The rule did not change; the world that produced its inputs did.

Common misunderstandings

• “Retraining on a schedule solves drift.” It helps only if recent data reflect the new situation and you can tell when performance actually changed.
• “Drift always shows up in the inputs.” The relationship can change while the inputs look the same.
• “It only affects machine learning.” Any fixed rule or threshold is exposed to it.