Problem
Pinterest shows ads on Home Feed (HF), Search (SR) and Related Pins (RP), and each surface had its own engagement (CTR) model. The three models started from one design but drifted apart in user sequence modeling, feature crossing, feature representations and training settings. A platform-wide improvement had to be rebuilt for each model, hyperparameters tuned on one surface did not transfer, and every idea was validated three times.
Example from the article: MMoE (multi-gate mixture-of-experts) on Home Feed alone, or a long-sequence Transformer on Search alone, gave no consistent gain.
Solution in brief
Merge the strongest components and features of the surface models and train on combined multi-surface data. Keep per-surface calibration, towers and checkpoint exports so each surface can still evolve. Offset the extra cost with projection layers and request-level embedding broadcasting. Unify cheap surfaces first, the expensive one (RP) later.
Questions this case answers
- How to consolidate several divergent per-segment prediction models into one shared model while keeping segment-specific quality, calibration and serving cost under control
- How to cut inference cost of a large ranking model when many candidates in one batch share the same user
- How to replace several per-surface CTR models with one model without losing quality on any surface?
- How to keep calibrated click probabilities when one model serves traffic from different surfaces?
Tools and methods
- Multi-gate mixture-of-experts (MMoE) — shared experts, one gate and tower per surface
- Long user sequence Transformer — encodes user history
- Surface-specific calibration layer — separate calibration of HF and SR clicks
- Multi-task learning with per-surface checkpoint exports — shared representation, one exported model per surface
- DCNv2 projection — shrinks Transformer outputs before crossing and tower layers
- Request-level broadcasting — one embedding lookup per unique user in a batch
Google or a one-time email code. No password.