The relationship between inputs and the target has changed, even though input distributions look stable. What is this?
Why B is the answer
Concept drift
Concept drift is a change in what the inputs mean for the outcome. Data drift is a change in the inputs.
In detail
- A changed relationship with stable inputs is concept drift.
- You need labels to confirm it.
Use case
An illustrative scenario, not a real incident.
Inputs look the same, but after a policy change the same customers behave differently. That is concept drift.
See it in different ways
flowchart TD A["Model issue"] --> B["Skew: training and serving data differ"] A --> C["Data drift: inputs change over time"] A --> D["Concept drift: the relationship changes"] A --> E["Attribution drift: feature importance shifts"]
mindmap
root((Decay))
Data drift
Inputs changed
Concept drift
Relationship changed
Skew
Training differs from serving
Attribution drift
Feature importance movedflowchart TD
Q{"Which moved?"}
Q -->|"The inputs"| A0["Data drift"]
Q -->|"The input to outcome relationship"| A1["Concept drift"]Key point: Concept drift changes what inputs mean.
Why the other options are wrong
- A. Data drift
- Data drift means inputs changed.
- C. Label leakage
- Leakage is a training problem.
- D. Quota exhaustion
- Quota is about capacity.