Data drift and model drift
When a model gets worse because the world changed, not the code.
Data drift means live inputs no longer look like the training data. Concept drift means the relationship the model learned has changed. Both degrade performance over time, so models need ongoing monitoring with thresholds, alerts and retraining decisions.
On the exam
Performance falling in production with unchanged code points to drift. The auditor checks whether monitoring should have caught it.
Related terms
- Model monitoringWatching a live model’s inputs, outputs and performance.
- Training, validation and test dataSeparate datasets to build, tune and fairly measure a model.
Test yourself with today’s question of the day.