concept drift

Concept drift occurs when the relationship between inputs and outputs changes over time, degrading model performance. **Definition**: P(Y|X) changes - same inputs now map to different outputs. The underlying patterns the model learned are no longer valid. **Example**: Customer buying behavior shifts due to economic changes, pandemic alters health data patterns, user preferences evolve. **Concept drift vs data drift**: Data drift is P(X) changing (input distribution). Concept drift is P(Y|X) changing (actual relationship). Both problematic. **Detection methods**: Monitor prediction accuracy with ground truth, statistical tests on residuals, track performance on labeled windows. **Types**: **Sudden**: Abrupt change (policy change, event). **Gradual**: Slow evolution over time. **Recurring**: Seasonal patterns. **Incremental**: Small continuous changes. **Response**: Retrain on recent data, use online learning, adaptive models, sliding window training. **Prevention**: Regular retraining schedules, continuous monitoring, domain expert alerts for known changes. **Challenges**: Ground truth delay makes detection slow, distinguishing drift from noise.