Problem
A pre-ranker scores many more candidates than the full ranker. Pinterest's first one was a two-tower model run separately on each retrieval source. Every new version had to be deployed to several services, auto-retraining needed all of them synchronized, and the final dot product barely captured user-item interaction or action sequences. A pre-ranker trained only on impressions also never sees the candidates it filters out at serving time.
Example from the article: the Pinterest home feed funnel, where the pre-ranking layer (Lightweight Scoring) sits between retrieval and the L2 full ranker.
Solution in brief
One model in a single cluster replaced the per-source rankers: a request-level user encoder plus an item-level scorer that crosses user and item features. Training data adds a logged sample of early-funnel candidates without impressions, and a distillation loss pulls scores toward the L2 ranker. Only this combination made a non-two-tower model win online.
Questions this case answers
- How to train a cheap early-stage ranker on data that matches the candidates it scores at serving time, when logged feedback exists only for items the later stages passed
- How to move from a late-fusion two-tower scorer to a model with early user-item feature crossing while keeping inference cost bounded
- How to replace a two-tower pre-ranking model with one that crosses user and item features early?
- How to build training data for a pre-ranking model that scores candidates the ranker never sees?
Tools and methods
- Request-level plus item-level model — user encoded once per request, item features and crossing computed online.
- Early funnel logging — logged candidates without impressions, joined with impression data.
- BCE with KL distillation — engagement labels plus alignment with calibrated L2 scores.
- Root-leaf inference service — items sharded by key so leaf hosts can cache their features.
Google or a one-time email code. No password.