Course contentsShow
← Machine Learning and Deep Learning
Lesson 3039 of 3,53866. Production Monitoring and Data DriftFree lesson

Understanding Concept Drift

What concept drift is and how the relationship between features and target changes over time in production systems.

Understanding Concept Drift

What you'll learn: How the statistical relationship between your features and target variable can silently change over time, breaking your model's predictions.

What Is Concept Drift?

You've learned about input drift (when P(X) changes) and output drift (when P(Y) changes). Concept drift is different and more insidious: it's when the fundamental relationship between inputs and outputs changes—when P(Y|X) shifts.

Think of it this way: your model learned that "cloudy skies" predict "rain" with 80% confidence. But due to climate change, that relationship shifts to 60%. The clouds look the same (no input drift), rain frequency is similar (no output drift), but the mapping between them has changed. Your model is now systematically wrong.

Why This Matters in Production

Concept drift is particularly dangerous because:

  • Your input monitoring won't catch it – feature distributions look normal
  • Basic output monitoring might miss it – prediction volumes seem fine
  • Only ground truth comparison reveals it – but that requires waiting for labels

Unlike covariate shift where you can retrain on new input distributions, concept drift means your entire learned hypothesis f(X) → Y is outdated. The world's rules have changed.

Real-World Examples

  • Credit scoring: Economic policy changes alter how income predicts default risk
  • Fraud detection: Fraudsters adapt tactics, so old patterns become obsolete
  • Medical diagnosis: Treatment protocols evolve, changing symptom-disease relationships
  • Recommendation systems: User preferences shift culturally (pandemic behavior changes)

Key Takeaway: Concept drift occurs when P(Y|X) changes over time, meaning the learned relationship between features and target becomes invalid even when input distributions remain stable—requiring model retraining with fresh labeled data to restore accuracy.