Авито устроило турнир uplift-моделей: победил X-learner, а не Causal Forest

Avito's trust and safety team analyst Grisha Kryukov held a "championship" in July 2026 among four popular uplift models — S-learner, T-learner, X-learner, and CausalForest — with a group stage and playoffs. The gold unexpectedly went to X-learner, which finished the group stage in only third place but won both semifinals and the final.

Why uplift models are needed

Uplift modeling estimates the individual effect of an intervention — how much a specific person's behavior will change because of a promotion, rather than the average response across a group. The main difficulty is that for a single object it's impossible to observe both outcomes at once: with the intervention and without it.

«The main problem: for a single player, we never know both situations at the same time», —

Grisha Kryukov, analyst on Avito's trust and safety team.

The four approaches solve this task differently: S-learner feeds the treatment flag in as an ordinary feature, T-learner trains two separate models, X-learner adds training on pseudo-effects and propensity weighting, while CausalForest builds trees that maximize the difference in effect.

How the tournament was structured

The tournament consisted of a group stage on synthetic data and playoffs on real data. The author measured model quality with the PEHE metric (Precision in Estimation of Heterogeneous Effect) — the lower it is, the more accurate the effect estimate.

• Models: S-learner, T-learner, X-learner, CausalForest
• Metric — PEHE, group points by ranking: 3-2-1-0
• Group stage — synthetic data, 5000 observations, 10 features, 70/30 split
• Semifinals — Hillstrom Mine That Data dataset, 64 thousand e-commerce customers
• Final — FIFA21 dataset, about 10 thousand footballers and 26 attributes

After three group-stage rounds, CausalForest led with 6 points, S-learner and X-learner each scored 5, and T-learner finished last with 2 points.

Why X-learner won

X-learner won because its architecture learns specifically from the difference between scenarios with and without the intervention, rather than from the overall response level. In the semifinal it crushed S-learner (PEHE 0.215 versus 0.747), and in the final it beat T-learner with a result of PEHE 0.579 versus 1.135.

«The architecture works as intended: instead of the overall level, the model learns specifically the difference between the worlds with and without the intervention», —

Grisha Kryukov.

The author explains the favorite's loss this way: on synthetic data, where the distribution is ideal, the advantage went to the robust CausalForest, but on "dirty" real-world data the more flexible architecture won out. T-learner suffered in particular: its two independent models don't exchange information, and this proved fatal in the final.

What this means

There is no universal winner in uplift modeling: on clean synthetic data and small samples CausalForest wins, while on noisy real-world data X-learner wins. The choice of model should be made based on the specific dataset, not on the method's reputation.

Frequently asked questions

What is uplift modeling?

It is a method for estimating the individual effect of an intervention — how much a specific user's behavior will change because of a promotion or an email, rather than the average effect across the whole audience. It's applied in marketing, retention, and anti-fraud.

Which model won the Avito tournament?

The champion was X-learner: it finished the group stage in only third place but won both semifinals and the final with a result of PEHE 0.579 versus 1.135 for T-learner. The reason is its architecture, which learns from the difference in effects.

What does the PEHE metric mean?

PEHE (Precision in Estimation of Heterogeneous Effect) measures how accurately a model estimates the individual effect: the lower the value, the closer the prediction is to the true one. It was used in the tournament to determine the standings.