Concept drift occurs when production data changes over time, causing a deployed model’s accuracy and relevance to decay. Machine learning recognizes patterns present in its training data, so predicting the future assumes that future behavior will resemble the past—a fragile assumption for user-generated and adversarial data. section 6.1.2
When yesterday’s data stops being representative
A music recommender trained in 2013 may be ineffective years later as listening behavior changes. Similarly, a sentiment model trained on IMDB reviews from 2011 may perform worse on reviews from 2020 because vocabulary, expressions, and movie genres evolve. Credit card fraud detection can face even faster drift, with fraud patterns changing practically every day. section 6.1.2
Treat drift as an ongoing maintenance problem rather than a one-time training defect. After launch, watch for changes in production data, collect and annotate new samples, improve the annotation pipeline, and focus especially on cases the current model finds difficult. Fast drift requires frequent retraining; expected model lifespans can range from days for fraud detection to weeks for recommendation systems or years for image search. section 6.3.4
