Problem
Pinterest predicted a daily notification budget for each user, once a day. That system could not react to new content or to what the user did during the day. The team needed near-real-time decisions on what to send, to whom, through which channel and when, across many notification types, with fewer irrelevant sends. A single score threshold drives the volume, so volume must stay on target while models change.
Example from the article: three candidate events are scored for push open, email click and unsubscribe. The top event per channel is sent only if its utility clears the threshold of the user's segment.
Solution in brief
Candidates are pre-generated daily into a per-user pool. A multi-head ranker scores them several times a day at the user's best delivery slots. A policy turns scores into a linear utility (engagement minus weighted unsubscribe risk) and sends when it beats a per-segment threshold. A PID-style controller tunes the thresholds to hold daily send volume.
Questions this case answers
- How to choose, for each user, which message to send, in which channel and when, so that engagement rises without raising opt-outs
- How to keep the volume of a threshold-based send policy on a target level when offline replay cannot mimic online streaming decisions
- How to decide which notification to send to a user, through which channel and at what time?
- How to keep the daily send volume on target when the send decision is a score threshold?
Tools and methods
- Event pool — per-user store of pre-generated candidates
- NEP ranker — multi-task model: push open, email click, unsubscribe
- Policy layer — utility and per-segment threshold pick channel and send or not
- PID threshold controller — holds send volume at target
- Random logging buckets — unbiased data for evaluation and policy tuning
- CUDA graphs — cut GPU serving overhead
Google or a one-time email code. No password.