DataTerminology.comNEW
Identifying changes in input data distributions that can degrade model performance and reliability over time.
DESCRIPTION
Data drift detection tracks distributional changes between training and live data. It uses statistical tests and metrics (e.g., PSI, KS test, KL divergence) to flag shifts. Detection informs retraining, feature updates, or rollback strategies. Monitoring integrates with model observability, capturing inputs, predictions, and outcomes. Addressing drift preserves accuracy and fairness, reducing model risk in production.
ALTERNATIVE DEFINITIONS
Evidently AI
Data drift is a change in the statistical properties of the input data, which can cause model performance to degrade.
Defines drift explicitly as a distributional change affecting performance.
AWS SageMaker
Amazon SageMaker Model Monitor detects deviations in data quality and model quality from training baselines and alerts you when they occur.
Operational framing with baseline comparisons and alerting.
Microsoft Learn
Dataset monitoring helps you identify data drift by comparing statistics between baseline datasets and new data over time.
Highlights baseline-versus-current statistical comparisons.
RELATED TERMS
Concept Drift Dataset Shift Model Monitoring Population Stability Index KL Divergence Feature Store Training-Serving Skew Shadow Deployment A/B Testing
COMPETING TERMS
Concept Drift Detection Change Point Detection
RELATIONS
- Broader: Drift Detection
KNOWLEDGE AREAS
MLOps AI Data Science Observability & Telemetry Testing & Validation
TARGET AUDIENCE
MLOps Engineer AI Engineer Data Scientist Chief AI Officer
EXTERNAL SOURCES
- Wikipedia - Concept drift
- AWS SageMaker - Model Monitor documentation
- Google Cloud - Vertex AI monitoring overview
- Microsoft - Dataset drift monitoring
TOP QUESTIONS
Why is Data Drift Detection critical in production ML?
It prevents silent performance degradation by signaling when live data diverges from training distributions.
Category: Relevancy
What metrics are commonly used for Data Drift Detection?
PSI, KL divergence, KS test, chi-squared, and Jensen–Shannon divergence are common choices.
Category: Overview
What are common mistakes in Data Drift Detection?
Using a single global metric, ignoring feature-level drift, or not setting actionable thresholds.
Category: Mistakes
How is Data Drift Detection operationalized?
Log inputs, compute periodic metrics, compare to baselines, alert on thresholds, and trigger retraining workflows.
Category: Management
What’s the risk of ignoring Data Drift Detection?
Accuracy declines, bias increases, and business decisions become unreliable, sometimes violating compliance.
Category: Risks
How does Data Drift differ from Concept Drift?
Data drift concerns input distributions; concept drift concerns the relationship between inputs and target.
Category: Opinion
What’s a hot topic around Data Drift Detection?
Unsupervised monitoring in LLM pipelines and synthetic data benchmarking.
Category: Topic
How to choose thresholds for drift alerts?
Use historical baselines, business impact mapping, and mixed severity levels for features and outcomes.
Category: Requirements
How can visualization aid Data Drift Detection?
Distribution plots, PSI heatmaps, and feature importance shifts help interpret and act on drift.
Category: Example