Championship Among Uplift Models: Testing S-learner, T-learner, X-learner and Causal Forest on Synthetic and Semi-Synthetic Data with PEHE Metric

The article, written by Grigory Kryukov, an analyst at Avito's trust and safety team, presents a championship among four uplift models: S-learner, T-learner, X-learner, and Causal Forest. Using a football World Cup analogy, the author describes each model's architecture: S-learner trains a single model with treatment as a feature; T-learner trains separate models for treatment and control groups; X-learner adds a second stage of pseudo-effects and a propensity model; Causal Forest adapts random forest to directly optimize for heterogeneous treatment effects. The tournament uses synthetic data (where true individual treatment effects are known) and semi-synthetic data to measure model quality via the PEHE metric. The group stage pits models against each other on synthetic data, followed by a knockout stage. The article concludes with results: Causal Forest wins the championship, being the most robust across scenarios, while X-learner shows strength on non-linear tasks but is complex. The full code is provided on GitHub.