Concept drift is what happens when the world changes the answer. A fraud model learns what fraud looked like last year; fraudsters adapt, and transactions that used to be safe now signal fraud. The inputs may look much the same, but the correct label for them has moved.
A supervised model learns a mapping from inputs X to outputs y. In probability terms it approximates P(y | X), the chance of each answer given the input. Lu and colleagues’ review breaks drift into sources. When only the input distribution P(X) changes, the decision boundary still holds; this is sometimes called virtual drift, and in production it is usually tracked as data drift. When P(y | X) itself changes, the boundary moves and accuracy falls. That is concept drift in the strict sense. Often both happen at once.
Shapes of drift
The same review describes drift by how it unfolds over time:
- Sudden: the concept switches abruptly, for example after a policy or product change.
- Gradual: the new concept increasingly replaces the old one over a period.
- Incremental: the concept slides steadily from old to new.
- Recurring: an earlier concept comes back, as with seasonal patterns.
Detecting and handling it
Because inputs can look normal, the most direct signal is the model’s own error rate on fresh labeled data, which often arrives late. Model monitoring therefore combines input statistics with delayed accuracy checks and proxy metrics. The usual responses are retraining on recent data, weighting recent examples more heavily, or online learning that updates the model continuously. The general term for the resulting drop in quality is model drift.