Concept Drift
Changes in the relationship between input features and target variables over time, causing model predictions to become less accurate.
In-Depth Explanation
Concept drift occurs when the underlying relationship between inputs and outputs changes, even if the input distribution stays the same. The "concept" the model learned no longer applies.
Concept drift types:
- Sudden: Abrupt change (e.g., new policy)
- Gradual: Slow transition over time
- Incremental: Step-by-step changes
- Recurring: Seasonal patterns
Examples:
- Customer preferences change
- Market conditions shift
- Regulations update
- Competitor actions
- Economic changes
Detection challenges:
- Need ground truth to detect (delayed feedback)
- Hard to distinguish from data drift
- May be subtle and gradual
Response strategies:
- Regular retraining schedules
- Performance monitoring
- Ensemble methods
- Online learning approaches
Business Context
Concept drift can make models obsolete even when they technically work correctly on new data. Regular validation against outcomes is essential.
How Clever Ops Uses This
We monitor for concept drift in Australian business ML systems, ensuring models are retrained when business dynamics change.
Example Use Case
"A credit risk model becoming less accurate as economic conditions change the relationship between customer attributes and default probability."