Problem
Design an online ad ranking system that predicts calibrated CTR and CVR to maximize auction eCPM
- Filter thousands of candidates under a strict fifty-millisecond latency budget
- Train separate click and conversion heads without gradient starvation
- Correct position bias and inject fresh inventory for new advertisers
Key Test Points
- Multi-stage pre-scoring
- MMOE multi-task learning
- Thompson sampling exploration
Spot an issue or have a suggestion?