Concept drift is a change over time in the relationship between a model’s inputs and the thing it predicts. The model learned that certain input patterns lead to certain outcomes, and the world has since changed those rules. A spam filter trained a year ago may still see similar-looking emails, but spammers have changed tactics, so messages once harmless are now spam. Because the correct answer for the same input has shifted, a model can lose accuracy even when the inputs look much as before [1]. Concept drift is one of the two main causes of model drift, alongside data drift, which concerns changes in inputs alone.
Definition
Formally, concept drift occurs when the conditional distribution of the target given the inputs, written P(Y given X), changes between training and deployment. Some authors distinguish real concept drift, where this relationship changes, from virtual drift, where only the input distribution changes but the decision boundary stays valid. In practice, both can occur together, and the terms are used loosely [1][2].
Examples
- Fraud and spam. Adversaries continually change tactics to evade detection.
- Consumer behavior. Tastes, price sensitivity and shopping channels evolve, and a pandemic or recession can overturn old patterns overnight.
- Credit risk. Economic conditions change what a given income and debt level means for default.
- Medicine. Disease strains, treatments and diagnostic criteria change.
- Language. Words acquire new meanings, so sentiment and topic classifiers age.
- Search and recommendation. User interests and the content available both change.
- Industrial systems. Equipment wears, so sensor readings that once signaled health now signal fault.
Types by how change happens
The literature groups concept drift by its pattern in time [1].
- Sudden (abrupt). One concept is replaced at once, such as after a new law or product launch.
- Gradual. The new concept appears alongside the old, with the share of the new rising, so data flips back and forth for a while.
- Incremental. The concept shifts through many small intermediate states.
- Recurring. Old concepts return, as with seasons or periodic events.
- Blips and noise. Short-lived anomalies that should be ignored.
The pattern matters because it determines what to do. Sudden change calls for fast replacement, while gradual change favors continuous updating, and recurring change rewards keeping old models for reuse.
Concept drift versus data drift
Data drift changes what the model sees. Concept drift changes what the right answer is. With data drift alone, a well-trained model that has seen enough variety may still be accurate, whereas concept drift always means that the learned mapping is out of date. Detecting concept drift typically requires knowing true outcomes, while data drift can be spotted from inputs alone.
Detection methods
- Performance-based. Track error rates over time and signal when they rise significantly. DDM (drift detection method) and EDDM watch the error rate of an online learner. Page-Hinkley and CUSUM detect shifts in the mean of a stream [1].
- Window-based. ADWIN keeps an adaptive window of recent data and shrinks it when it detects that older and newer parts differ [1].
- Distribution-based. Compare feature and prediction distributions between windows, with the caution that these can miss concept changes that leave inputs unchanged.
- Model-based. Monitor disagreement between a long-term and a short-term model, or between ensemble members.
- Delayed-label strategies. When labels arrive late, use proxies such as business metrics, human review samples or user corrections, then confirm when ground truth arrives.
- Unsupervised signals. Shifts in prediction confidence or in the rate of positive predictions.
Adaptation strategies
- Periodic retraining on a recent window of data, the simplest and most common method.
- Triggered retraining when a detector fires.
- Online and incremental learning, updating the model with each new example or small batch; see model deployment.
- Instance weighting and forgetting. Give recent examples more weight or discard old ones.
- Ensembles. Keep several models trained on different periods and weight them by recent accuracy, so that the best suited model wins as conditions change.
- Meta-learning and fine-tuning to adapt quickly with few examples; see transfer learning.
- Active learning. Request labels for the most informative recent examples; see active learning.
- Model redesign with new features when the concept has fundamentally changed.
The central tension is the stability-plasticity dilemma: adapting too quickly causes the model to chase noise and forget useful knowledge, whereas adapting too slowly leaves it stale.
Concept drift in practice
Production teams monitor outcome metrics by segment, keep a rolling evaluation set built from recent labeled data, version every model and dataset, and rehearse rollbacks. They separate pipeline errors from true drift before retraining, because a feed that silently changed units looks like drift but needs a fix, not a new model. See AI observability and LLMOps.
Concept drift and large language models
For language models, the “concept” can shift in several ways: facts change after the training cutoff, social norms and acceptable content evolve, and user expectations and tasks change. Retrieval of current documents, regular evaluation on fresh questions and periodic updating of models and prompts address these; see retrieval-augmented generation.
A worked scenario
A streaming service trains a model to predict which new shows a viewer will finish. It works well for months. Then the service changes its autoplay behavior, a popular competitor leaves the market and a holiday season shifts viewing habits. The same viewer profile that once predicted binge-watching of dramas now predicts short comedy sessions. Input features look similar, so data drift checks stay quiet, but the weekly measure of how many recommended shows are completed starts to fall. A detector on the completion rate raises a flag. The team compares the recent window with older ones, confirms that the relationship has changed for several user groups, and trains a new model on the most recent three months, using older data at reduced weight. They deploy it first to a small share of users and expand once completion rates recover.
Why concept drift is hard
Several things make it harder than other monitoring problems. True outcomes are often delayed, so evidence of drift arrives late. Drift can hide in one segment while the overall average stays stable. Detection methods trade off false alarms against delay: sensitive detectors fire on noise, while cautious ones miss real change. Adaptation can itself create feedback loops, since a model that changes what it recommends changes the behavior it will later learn from. And adversarial drift, where opponents react to the model, can turn the learning problem into a contest in which no fixed model remains correct for long.
Planning for change
Since drift is expected, mature teams design for it from the start. They choose features that are stable and causal where possible, preferring signals that reflect underlying mechanisms to those that merely correlate. They keep training pipelines automated so that retraining is routine, hold back recent labeled data for validation under current conditions and set explicit service levels for how fast the system must detect and respond. Documentation lists which assumptions the model relies on, so that when the world changes, the team can quickly see which ones broke. See also AI governance for oversight of models in production.
FAQs
What is concept drift? It is a change over time in the relationship between a model’s inputs and the target it predicts.
What are the types of concept drift? Sudden, gradual, incremental and recurring.
How do you detect concept drift? By monitoring error rates and using detectors such as DDM, ADWIN and Page-Hinkley, plus checks when labels are delayed.
What is the difference between concept drift and data drift? Concept drift changes the mapping from inputs to outputs, while data drift changes only the inputs.
How do you handle concept drift? Retrain on recent data, use online learning or ensembles, and redesign features when needed.
More in AI Evaluation and Reliability
Key facts
- Concept drift is a change in the relationship between inputs and the target variable over time, so a model that was accurate becomes inaccurate even if the inputs look similar1.
- Drift is commonly categorized as sudden, gradual, incremental or recurring, depending on how the new concept replaces the old1.
- Detection methods for data streams include DDM, EDDM, ADWIN and Page-Hinkley, which monitor error rates or statistics to signal change1.
Related concepts
Quick checkWhat is concept drift?Show answer
Concept drift is a change over time in the statistical relationship between a model's inputs and the target it predicts, so that patterns learned during training no longer hold and the model's accuracy declines.
Sources
- Joao Gama et al., A Survey on Concept Drift Adaptation, ACM Computing Surveys, https://doi.org/10.1145/2523813
- Jie Lu et al., Learning under Concept Drift: A Review, IEEE Transactions on Knowledge and Data Engineering, https://doi.org/10.1109/TKDE.2018.2876857
Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue
