Causal Inference and Uplift Modeling for Intelligent Subsidy in Hotel Marketing at Hello Mobility

Background : Hello Mobility offers two‑wheel shared rides and hotel services. To maximize total utility in hotel promotions, the company issues coupons as subsidies, aiming to convert users who would not purchase otherwise.

Intelligent Subsidy Analysis : Users are segmented into four quadrants (marketing‑sensitive, naturally converting, indifferent, and adverse) based on coupon receipt and purchase behavior. The goal is to target the marketing‑sensitive group to improve subsidy efficiency.

Limitations of Traditional Response Models : Conventional CTR‑based response models predict purchase probability but cannot capture the causal effect of issuing a coupon, leading to sub‑optimal subsidy strategies.

Causal Inference & Uplift Modeling : The article introduces uplift (incremental) modeling, which estimates the change in conversion caused by an intervention. It describes three meta‑learners—T‑Learner, S‑Learner, and X‑Learner—and notes direct uplift methods such as tree‑based models and deep‑learning approaches like DragonNet.

Uplift Evaluation : Because true uplift labels are unavailable, the offline metric AUUC (Area Under the Uplift Curve) is used. The evaluation pipeline includes scoring, sorting by uplift score, bucketing, computing cumulative gain, and integrating the curve.

Tree‑Based Uplift Model (Treelift) : To align model objectives with business goals, the split criterion is changed from KL‑divergence to the squared difference of per‑user utility between treatment and control groups. The algorithm recursively selects the feature that maximizes this utility gain, producing the Treelift model.

Real‑Time Decision System : The online system routes traffic to three groups: a manual rule‑based group, a small‑scale random‑experiment group (providing training data), and an algorithmic group where the Treelift model decides coupon issuance.

Results : Offline tests show Treelift achieves the highest AUUC among T‑Learner, S‑Learner, and other baselines. Online A/B tests report a 4.7% lift over manual strategies and a 2.3% lift over the previous response model.

Future Work : Plans include reducing training latency, exploring pruning and regularization, applying propensity‑score matching to leverage observational data, and extending the framework to scenarios with explicit cost constraints using integer programming.