How We Solve Yandex Lavka's Optimization Problems with Uplift Models

Vyacheslav Kostrov23 May 2026

Problem

A delivery app wants to grow turnover (GMV) with delivery discounts, but the budget is limited and discounts cannot go to everyone. Every extra ruble of turnover is paid for with discount volume or lost profit. A model that ranks users only by predicted turnover gain sends money to the wrong people.

Example from the article: user A brings +200 rubles of turnover, user B +120. For A the company paid 110 rubles, for B only 20, so B gives the better net result.

Solution in brief

Train uplift models on randomized-experiment data and build the cost into the model. Either put the cost into the target through a Lagrangian with a tuned λ, or, with uplift trees, make the split criterion the turnover-per-cost trade-off. Evaluate the final policy on the randomized data before any A/B test.

Questions this case answers
  • How to choose which units get a costly treatment under a budget constraint when the benefit must be measured net of the cost of the treatment
  • How to choose among several treatment levels per unit using randomized historical data before a live test
  • How to choose which users get a discount or promo under a budget constraint
  • How to build an uplift model that accounts for the cost of the treatment

Tools and methods

  • Uplift trees and uplift forest — split by difference in treatment effect
  • Conditional average treatment effect (CATE) — uplift estimate per user
  • Uplift curve with a cost axis — ranking quality with cost on the X axis
  • Lagrangian constraint — budget limit inside the target, parameter λ
  • Custom split criterion — trees separate segments by trade-off, no λ
  • Policy evaluation on randomized data — matches predicted and actual treatment
Continue reading — it's free

Google or a one-time email code. No password.