Problem
Yandex Plus is a paid subscription. Its Plus Daily feature gives subscribers a gift for completing tasks: a bundle of about 10 promo codes for other Yandex services. Each code brings a new user to a service, and that service pays Plus a reward. The team must decide per user how long the task progress bar is and how the codes are ordered. Fixed business logic treats all users alike. New codes also appear every few weeks, and at first only claiming of a code was logged, not its use.
Example from the article: a current Market promo code gives 700 rubles off a purchase from 6000 rubles. In the speaker's profit example, acquiring a Market user costs 2000 rubles and the code costs 500.
Solution in brief
Predict the probability that a user applies a given promo code, then rank codes by expected profit: reward times that probability. The same probabilities decide whether to shorten or lengthen the progress bar. A single CatBoost classifier takes promo code service and parameters as features, so a new code needs no new model.
Questions this case answers
- How to rank a bundle of offers for each user by expected profit, when the probability of use must be estimated for offers that keep appearing
- How to set the length of a per-user task progress bar that drives completion of a reward, when the completion event is only partly logged
- How to rank promo codes in a reward bundle so that the company earns the most from user acquisition
- How to choose the length of a task progress bar per user to push them to claim and apply a reward
Tools and methods
- Binary classifier (S-learner style) — probability that a promo code is applied, given user, service and code parameters.
- CatBoost with LogLoss — the model and its loss.
- Reward × score ranking — orders the codes in a bundle.
- Positive-to-negative score ratio — sets the progress bar length.
Google or a one-time email code. No password.
