DataTerminology.comNEW
Identifying shifts in data distributions or model performance that indicate changing real-world conditions.
DESCRIPTION
Drift detection is the process of monitoring for statistically significant changes in input data, model features, or predictions after deployment. It helps detect concept drift, data drift, and performance degradation. Typical techniques include statistical tests, divergence measures, PSI, and change point detection. It triggers alerts, root-cause analysis, and retraining workflows to maintain model reliability and compliance. Drift detection is a core capability of model monitoring within MLOps.
ALTERNATIVE DEFINITIONS
Google Cloud
Model monitoring detects skew and drift in your data and predictions to help you maintain model quality in production.
Emphasizes production model quality and monitoring for both data and prediction drift.
Microsoft Learn
Data drift monitoring measures changes in data over time to identify when your model’s training data no longer represents production data.
Highlights training vs. production mismatch as the key risk detected.
Seldon Alibi Detect
Drift detection aims to detect a change in the distribution of data over time, indicating that the current model may no longer be valid for the new data.
Connects detected distribution change directly to model validity concerns.
RELATED TERMS
Concept Drift Data Drift Model Drift Domain Shift Population Stability Index Model Monitoring KL Divergence A/B Testing Retraining
COMPETING TERMS
Change Point Detection Data Anomaly Detection
RELATIONS
- Narrower: Change Point Detection Data Drift Detection
KNOWLEDGE AREAS
MLOps Observability & Telemetry AI Data Science Testing & Validation
TARGET AUDIENCE
MLOps Engineer Data Scientist AI Engineer Data Engineer Chief AI Officer Data Architect
EXTERNAL SOURCES
- Wikipedia - Concept drift
- Google Cloud - Vertex AI model monitoring overview
- Microsoft - Azure ML data drift monitoring guide
- Seldon - Open-source library for drift detection
- Evidently AI - Practical guides on data and model drift
TOP QUESTIONS
What is Drift Detection in machine learning?
Drift detection identifies changes in data distributions or model performance over time. It ensures deployed models remain accurate as real-world conditions evolve.
Category: Definition
Why is Drift Detection important for production ML systems?
Without drift detection, models silently degrade, causing poor predictions and business risk. Detecting drift triggers investigation, retraining, or rollback to maintain reliability.
Category: Relevancy
What types of drift does Drift Detection cover?
It commonly covers data drift, concept drift, covariate shift, and prediction drift. Many teams also track performance drift, like accuracy or AUC decline.
Category: Overview
How do you implement Drift Detection in practice?
Use statistical tests and divergence metrics (e.g., PSI, KL, JS), set thresholds, monitor in batch or streaming, and alert when violations occur.
Category: How-to
What tools support Drift Detection?
Popular options include Vertex AI Model Monitoring, Azure ML data drift, Seldon Alibi Detect, Evidently AI, and custom pipelines on Spark or Pandas.
Category: Tooling
What is a common mistake in Drift Detection?
Using arbitrary thresholds without baselining leads to noisy alerts. Calibrate thresholds and segment by population to avoid false positives.
Category: Mistakes
How often should Drift Detection run?
Cadence depends on data velocity and risk. Many teams use daily jobs for batch systems and near-real-time checks for streaming or critical applications.
Category: Operations
How is Drift Detection different from anomaly detection?
Anomaly detection flags rare events, while drift detection focuses on persistent distribution shifts. They complement each other in robust monitoring.
Category: Comparison
What happens after Drift Detection triggers an alert?
Investigate root causes, assess impact, and decide actions like feature fixes, data pipeline corrections, or model retraining and redeployment.
Category: Process