Your model works well on validation but fails in a new geography, product surface, or customer segment. How do you reason about domain shift?
formulate your answer, then —
tldr
Domain shift means deployment differs from training. Diagnose whether inputs, priors, or P(y|x) changed. Use deployment-like validation splits, target-domain data, segment monitoring, recalibration, fine-tuning, and fallback behavior.
follow-up
- How do covariate shift, label shift, and concept drift differ?
- How would you design a validation split for launching in a new country?
- What should a system do when it detects OOD inputs?