Data drift
- What
- Feature
- Population
- Covariate shift
- How
- Enough of the data needs to be labeled to introduce new classes
- Retrain the model
Concept drift
- What
- Pt(X, y) != Pt+1(X, y)
- Sudden
- Gradual
- Incremental
- Recurring
- How
- The old data needs to be relabeled. Retrain the model
- Use an ensemble approach to train your new model
Prediction drift
Label drift
Training, serving skew
- Tensorflow data validation
