DataTerminology.comNEW
Changes over time in the relationship between inputs and outputs, degrading model performance unless detected and addressed.
DESCRIPTION
Concept drift arises when the underlying mapping from features to target evolves, causing models to lose accuracy. It is common in dynamic domains such as finance, operations, and user behavior. Drift detection monitors performance, prediction distributions, and error patterns to trigger retraining or recalibration. Techniques include sliding windows, drift detectors, and online learning. Continuous monitoring and governance processes are critical for resilience in production ML.
ALTERNATIVE DEFINITIONS
Wikipedia
In machine learning and data mining, concept drift means that the statistical properties of the target variable, which the model is trying to predict, change over time.
Focuses on changes in the target’s statistical properties over time.
NannyML
Concept drift occurs when the relationship between features and the target changes over time, making the model’s learned patterns outdated.
Highlights real-world monitoring implications and corrective actions.
RELATED TERMS
Dataset Shift Domain Shift Model Monitoring Drift Detection Model Retraining Stationarity MLOps
COMPETING TERMS
KNOWLEDGE AREAS
AI Data Science MLOps Observability & Telemetry
TARGET AUDIENCE
Data Scientist MLOps Engineer AI Engineer Data Strategy Consultant
EXTERNAL SOURCES
- Wikipedia - Concept drift
- scikit-multiflow - Streaming and drift detection toolkit
- Evidently AI - Drift overview and detection
- Seldon - Production drift strategies
TOP QUESTIONS
What causes Concept Drift?
Shifts in user behavior, market conditions, policies, or external environments can alter the input-output relationship and degrade model accuracy.
Category: Example
How is Concept Drift detected?
Monitor performance metrics over time, use drift detectors, and compare prediction distributions with historical baselines.
Category: Requirements
How often should models be retrained to handle Concept Drift?
Retrain based on drift signals, performance thresholds, or business cycles. Overly frequent retraining can be wasteful; too infrequent risks inaccuracy.
Category: Opinion
What is the difference between Concept Drift and Data Drift?
Concept Drift changes the target relationship; Data Drift changes input distributions. Both affect performance but require different detection methods.
Category: Misunderstanding
Can online learning mitigate Concept Drift?
Yes, online and incremental learning adapt continuously, improving resilience to evolving data.
Category: Requirements
How is Concept Drift managed in MLOps?
Use monitoring pipelines, alerting, retraining policies, model versioning, and shadow deployments to validate updates safely.
Category: Management
What is a hot topic around Concept Drift?
Robust drift detection under limited labels and explainable alarms to aid human oversight is actively researched.
Category: Topic
How does Concept Drift impact fairness?
Shifts can amplify biases if models fail to recalibrate, so fairness monitoring is essential alongside performance monitoring.
Category: Risk
Where to start learning about Concept Drift?
Explore scikit-multiflow for detectors and vendor blogs illustrating practical pipelines and metrics.
Category: Getting started