A change in the relationship between the inputs and the target, P(y | X), so the same inputs now call for a different prediction. The model's learned rules go stale even when the input data looks unchanged.
Card 1 of 228. Answer: A change in the relationship between the inputs and the target, P(y | X), so the same inputs now call for a different prediction. The model's learned rules go stale even when the input data looks unchanged.
About · Quizzes · History of AI · To suggest new questions or report an error send me a dm.
Concept drift is a change in P(y | X), the mapping the model learned. Unlike data drift, the inputs can look exactly as they did in training; what has changed is what they mean for the target. It can be sudden (a new regulation), gradual (tastes shifting over months) or recurring (seasonal patterns).
Example: a fraud model learned that one large gift card purchase is risky and small ones are safe. Fraudsters notice and switch to many $50 purchases. The input distribution barely moves, so a data drift monitor stays quiet, but a $50 gift card purchase that used to be almost always legitimate is now often fraud. The model keeps calling it safe, and recall falls.
Because the inputs may not move, concept drift shows up in the labels: monitor realized performance once ground truth arrives, or compare the live model with one retrained on recent data. The usual fix is retraining on recent data, with a sliding window or weights that favour new examples. Interviewers often want the contrast in one line: data drift changes P(X), concept drift changes P(y | X), and label shift changes P(y).
