Problem
Uber Eats marketing, city operations and product teams all send pushes, and volume reached billions per month by the end of 2020. Teams sent conflicting messages minutes apart, pushes went out after hours or with invalid deeplinks and promo codes, and users got little personalization of content, time or frequency. Marketing teams also spent upwards of 15 hours per week per team member on manual orchestration to avoid conflicts.
Example from the article: a user's inbox holds three pushes, the cap is 2 pushes per day over a 7-day horizon, and each push has an expiry date. The system has to choose which pushes to send, and on which day and hour.
Solution in brief
All pushes go into a per-user inbox instead of being sent directly. Each time one arrives, a scheduler scores every (push, time) pair with an XGBoost model that predicts an order within 24 hours. An integer linear program then picks the pairs with the highest total score under business constraints.
Questions this case answers
- How to pick, for each user, which queued messages to deliver and at what times so that a predicted-response objective is maximized under frequency, expiry and time-window constraints
- How to turn a per-item response probability model into a global schedule instead of greedy per-message decisions
- How to choose which buffered push notifications to send to a user, and when, to maximize orders without spamming?
- How to enforce frequency caps, expiration dates and send windows while still using a conversion model?
Tools and methods
- XGBoost — predicts the probability of an order within 24 hours for a (user, push, time) triple.
- Integer linear programming (assignment problem) — picks the schedule with the highest summed score under caps, windows and expiry.
- Downsampling of the negative class and feature pruning — handle the rare-conversion training set.
- Apache Kafka — buffers schedule requests, per-hour send queues and delivery.
Google or a one-time email code. No password.