DataTerminology.comNEW
A change in input data distributions over time that can degrade model performance.
DESCRIPTION
Data drift occurs when the statistical properties of input features shift between training and production. Causes include population changes, upstream pipeline modifications, and data quality issues. Detection uses statistical tests, distance metrics, and monitoring dashboards. Mitigation includes retraining, feature engineering, and robust preprocessing. It is distinct from concept drift, which pertains to the target relationship. Proactive monitoring reduces business risk.
ALTERNATIVE DEFINITIONS
Wikipedia
Covariate shift is a special case of dataset shift where the distribution of the input variables changes but the conditional distribution of the output given the inputs remains the same.
Clarifies data drift as input distribution changes (P(X)).
Google Cloud
Data drift occurs when the statistical properties of the input data change over time.
Concise input-focused drift definition within MLOps guidance.
RELATED TERMS
Domain Shift Dataset Shift Population Stability Index KS Test Monitoring Retraining Feature Distribution Model Drift Drift Detection
COMPETING TERMS
KNOWLEDGE AREAS
AI MLOps Observability & Telemetry Testing & Validation
TARGET AUDIENCE
MLOps Engineer AI Engineer Data Scientist Data Steward
EXTERNAL SOURCES
- Wikipedia - Dataset shift
- Microsoft Learn - Monitor dataset drift
- scikit-learn - Related data issues
TOP QUESTIONS
What causes Data Drift in production?
Shifts in user behavior, seasonality, new markets, or upstream data processing changes can alter feature distributions.
Category: Overview
How do I detect Data Drift?
Use statistics like PSI, KS tests, population means, and visualization of feature histograms and quantiles.
Category: Monitoring
How is Data Drift different from Concept Drift?
Data drift affects P(X); concept drift affects P(Y|X). Both can degrade performance.
Category: Comparison
What are common mistakes when addressing Data Drift?
Overfitting to recent noise or ignoring data quality regressions that masquerade as drift.
Category: Mistakes
How frequently should I monitor for Data Drift?
Match monitoring cadence to data velocity and risk—real-time for streaming, daily or weekly for batch.
Category: Best practices
What mitigations help with Data Drift?
Retrain models, recalibrate probabilities, engineer stable features, and fix upstream quality issues.
Category: Mitigation
Can robust models eliminate Data Drift risk?
No. Robustness reduces sensitivity, but real-world distributions change; monitoring remains essential.
Category: Risk
How do I prioritize features for drift monitoring?
Focus on features with high importance, known instability, or critical business impact.
Category: Guidance
What’s an example of Data Drift?
A credit model’s income distribution shifts after policy changes, impacting default risk predictions.
Category: Example