Problem
Yandex Lavka releases most courier shifts once a week: analysts compute the load per darkstore (where orders are assembled and handed to couriers), cut slots, and all shifts appear in the app at once. Couriers stop working for about half an hour to sign up shift by shift, and they over-book for fear that others will take the slots, then drop out. If fewer couriers show up than planned, the service raises delivery prices (surge) to suppress demand.
Example from the article: 12 hours of Friday slots appear at once, and a courier signs up for each shift separately.
Solution in brief
Compute a week of shift recommendations per courier from his last five weeks of work history and send them as a push notification half an hour before the general release. Shifts are distributed greedily, highest-rated couriers first, and each assigned shift leaves the shared pool.
Questions this case answers
- How to distribute a limited shared pool of slots among many agents so that no slot is offered to two agents at once
- How to build a first recommender from a short behavioural history without training a model
- How to recommend work shifts to gig couriers so that shifts get filled and couriers keep their hours
- How to recommend items that compete for a limited resource, so that two users are not shown the same slot
Tools and methods
- 168-hour preference heatmap — per-courier vector of the probability to take and work each hour of the week
- Dot product with a shift vector — estimates the probability that the courier shows up for a shift
- Greedy assignment by courier rating — removes assigned shifts from the pool
- Runtime recommendations on button tap — second pipeline for mid-week shifts and new couriers
- Ranking model (planned) — binary classifier of shift acceptance
Google or a one-time email code. No password.
