DataTerminology.comNEW
A change over time in the distribution of input features that can degrade model performance.
DESCRIPTION
Feature drift occurs when the statistical properties of input variables shift from training to production. Causes include seasonal patterns, new user behavior, and data pipeline changes. Monitoring involves distribution tests, PSI, and KS statistics. Detecting drift prompts retraining, data quality checks, or model recalibration. It is distinct from concept drift, which concerns the target relationship changing.
ALTERNATIVE DEFINITIONS
Microsoft Learn
Data drift is the change in model input data that leads to model performance degradation over time.
Defines drift as input change linked to performance impact, applicable to feature drift.
AWS
Model Monitor can detect data drift by comparing statistics from baseline datasets to real-time production data to identify significant deviations.
Operational perspective on detecting drift via statistical deviation monitoring.
Evidently AI
Data drift refers to changes in the distribution of input features; concept drift is a change in the relationship between inputs and the target.
Clear differentiation between feature (data) drift and concept drift.
RELATED TERMS
Data Drift Domain Shift Population Stability Index Model Monitoring MLOps
COMPETING TERMS
Concept Drift Label Drift
KNOWLEDGE AREAS
MLOps Observability & Telemetry Testing & Validation AI Data Science
TARGET AUDIENCE
MLOps Engineer AI Engineer Data Scientist Data Steward IT Manager
EXTERNAL SOURCES
- Wikipedia - Concept drift
- Microsoft - Azure guide for data drift monitoring
- AWS - Detect drift using SageMaker Model Monitor
- Evidently AI - Explainer comparing data and concept drift
TOP QUESTIONS
How do I detect Feature Drift in production?
Compare feature distributions to a baseline using PSI, KS tests, or distance metrics. Automate checks and alerts in your monitoring pipeline.
Category: Testing
What causes Feature Drift most often?
Seasonality, new user cohorts, upstream schema changes, and data collection bugs commonly shift feature distributions.
Category: Examples
How is Feature Drift different from Concept Drift?
Feature drift changes input distributions; concept drift changes the input–target relationship. Both degrade performance but need different responses.
Category: Comparison
What should I do when Feature Drift is detected?
Investigate upstream changes, retrain with recent data, recalibrate thresholds, and update feature engineering as needed.
Category: Mitigation
How often should I evaluate Feature Drift?
Continuously for streaming and at least daily or weekly for batch. Adjust cadence to business risk and data velocity.
Category: Best practices
What is a common mistake with Feature Drift monitoring?
Using a single global threshold for all features. Prioritize by business impact and sensitivity; tune thresholds per feature.
Category: Mistakes
Can Feature Drift be beneficial?
It can reveal new market segments or product usage. Treat it as a signal to adapt models and products.
Category: Opinion
How does Feature Drift affect retraining schedules?
Frequent drift may require more frequent retraining or online learning. Balance freshness with cost and stability.
Category: Management
Which features should be monitored for drift?
Monitor all critical inputs and derived features, especially those with high model importance or compliance implications.
Category: Requirements