For time-ordered support tickets used to evaluate drift-sensitive behavior, how should train, validation, and test data usually be split?
- ARandomly across all dates with no regard for time
- BBy time order so later data simulates future use
- CBy ticket length only
- DBy the model's predicted confidence