← AI Terminology
Concept Drift
Concept drift is when the relationship between inputs and the target changes over time (P(y|x) shifts), even if input margins look similar.
Harder than pure covariate drift because labels’ meanings move.
Harder than pure covariate drift because labels’ meanings move.
Why It Matters in AI
Fraud patterns evolve; user intent for the same words changes; policies redefine “allowed.” Models need label-aware monitoring and faster adaptation loops. Critical in finance, content moderation, and ads.
Key Points
| Aspect | Description |
|---|---|
| LLM | Policy changes; factual world changes |
| Detect | Performance drop with delayed labels; challenger models |
| Related | Data drift, continual learning |
| Mitigate | Online learning, frequent retrain, human review |
| Definition | Posterior P(y |
| Vs data drift | Inputs may look fine while labels change |
Simple Analogy
The phrase “that’s sick” flipping meaning across generations — same words, new label of positive vs negative.
Common Usage Examples
- Track precision/recall with delayed ground truth
- Shadow new policy classifiers
- Short-lived models for volatile targets
- Human audit samples weekly
Summary
In short: Concept drift is when the input-to-label mapping itself changes — models must relearn meaning, not only new input styles.