Problem
Yandex Go (Yandex.Taxi) has user tasks (choosing communications, segmentation) and driver tasks (load prediction, churn prevention). Each needs model features. Aggregates over ride history (order count, average discount, payment method shares) are easy to explain, but they often miss the properties that matter most. Proving with an online experiment that each new model beats the old one is expensive.
Example from the article: a lookalike model had to find users likely to take the "Kids" tariff for the first time next month, starting from a set of past first-time riders and random users.
Solution in brief
Treat a user's history as a sequence of events, train a transformer encoder to predict the second part of the history from the first, and average its last layer to get one universal embedding. Compare embeddings with aggregate features offline on uplift curves from a single past experiment. Explain embedding components with Integrated Gradients and Shapley values.
Questions this case answers
- How to learn a universal representation of an entity from its event sequence when hand-built aggregate features miss important behavior
- How to compare candidate models for a targeting task from a single randomized experiment instead of one online test per model
- How to turn a user's event history into a reusable vector instead of hand-built aggregates?
- How to find lookalike users for a target audience from ride history?
Tools and methods
- Transformer encoder (encoder-only, last layer averaged) — turns a sequence of ride "transactions" into a user or driver embedding.
- Uplift curve — ranks users by model score against one randomized test/control experiment, so many models can be compared without new online tests.
- Integrated Gradients and Shapley values — show which input features drive each embedding component.
- Model registry and monthly batch inference — keep training reproducible and embeddings consistent over time.
Google or a one-time email code. No password.
