Concept Drift vs Data Drift (and Why Most Teams Detect Neither)
Published
•1 min readThese two are often confused.
They are not the same.
Data Drift
Input distributions change.
Example:
User age distribution shifts.
Concept Drift
The relationship between input and outcome changes.
Example:
Users stop clicking what they used to.
This is far more dangerous.
Why Concept Drift Is Hard
You only detect it when labels arrive.
Often weeks later.
By then, damage is done.
Detection Techniques
Prediction confidence decay
Performance on delayed labels
Statistical tests on outcomes
Drift detection must be continuous.
Final Thought
Data drift changes inputs.
Concept drift changes reality.
Ignore either at your own risk.
