The job of ranking
The ranking stage takes the few hundred candidates from retrieval and assigns each a score that predicts how much the user will engage. It then sorts by score. Because the set is small, it can afford a deep model with rich features.
What the model predicts
- A common target is probability of a positive action, such as a click, watch, or purchase.
- Many systems predict several actions at once, like click probability and dwell time.
- These predictions feed a final score used for ordering.
Features that matter
- User features: history, demographics, recent context.
- Item features: category, age, quality signals.
- Cross features: interactions between the specific user and item, often the strongest signal.
Training the ranker
Rankers learn from logged interactions: shown items that were engaged become positives, ignored ones become negatives. A key subtlety is that the model only ever sees items retrieval surfaced, so its view of the world is shaped by upstream stages.
Latency and depth
Ranking depth is bounded by latency. A richer model raises quality but costs milliseconds per candidate, so teams tune model size against the request budget.
Key idea
The ranking stage applies a heavy feature rich model to a small candidate set, predicting engagement probabilities and sorting by them within a strict latency budget.