How do you handle cold-start users / items in production?
mediumAnswer
- (1) Content-based fallback: use features (demographics, item metadata) via a separate model.
- (2) Global popular items.
- (3) Simple heuristics (category-popular).
- (4) Bandits / Thompson sampling to explore fast.
- (5) Two-tower model that generalizes via features.
- (6) Hybrid: gradually blend into personalized as history accumulates.
- Cold-start is not a modeling failure — it's a system design pattern (product should still function).
Check yourself — multiple choice
- No solution
- Content-based fallback + global popular + heuristics + bandits + two-tower feature generalization + gradual blend; system pattern (product still functions)
- Random
- Wait for data
Cold-start: content fallback + popular + bandits + two-tower + blend.
#deployment#features
Practise MLOps & Data Quality
215 interview questions in this topic.