Tree‑Based Causal Inference for Smart Subsidy Optimization at Hello Mobility

The talk introduces the problem of intelligent subsidy in Hello Mobility's hotel marketing, where coupons are issued to increase total utility, but traditional response models fail to capture the true incremental effect of the intervention.

It then explains causal inference and uplift modeling, distinguishing correlation‑based response prediction from true uplift estimation, and describes common uplift learners such as T‑Learner, S‑Learner, and X‑Learner, as well as direct uplift models like tree‑based methods and deep‑learning approaches (e.g., DragonNet).

The core contribution is a customized tree‑based uplift model (Treelift) that aligns the split criterion with the business goal of maximizing per‑user utility rather than using generic KL‑divergence; the model evaluates split gain by the squared difference of average utilities between treatment and control groups.

Model training relies on small‑scale random experiments to obtain unbiased treatment‑control data, and offline performance is measured with AUUC (area under the uplift curve), which reflects how well the model ranks users by true incremental conversion.

Online deployment separates traffic into manual rules, random‑experiment groups, and algorithmic decisions, with the Treelift model driving coupon issuance; real‑world A/B tests show a 4.7% lift over manual strategies and a 2.3% lift over the previous response model.

Future work includes reducing training latency, exploring pruning and regularization, applying propensity‑score matching to leverage observational data, and extending the framework to other domains where cost considerations require more sophisticated optimization.