Applying Uplift Modeling, PSM Matching, and Spark CausalML for Growth at Tencent Video

The concept of causal inference has become a hot topic in the past two years, with many business scenarios emerging that align closely with growth objectives. Tencent Video’s growth team has experimented with causal inference methods and shares three main parts: uplift gain models, PSM user matching, and Spark CausalML.

Uplift Gain Model defines the intervention as coupon issuance and classifies users into four quadrants (always‑convert, opposite‑convert, uplift‑sensitive, never‑convert). By comparing conversion rates with and without the coupon, the uplift model more accurately identifies users who generate incremental growth, guiding fine‑grained marketing strategies.

The model comparison includes S‑learner, T‑learner, X‑learner, and DragonNet, with evaluation using uplift quantile plots and Qini curves rather than traditional RMSE.

Application Cases for Uplift : Ad placement: targeting the top 30% of users with highest uplift improves ROI and playback conversion. Internal flow positions: modeling the causal effect on next‑day retention helps filter out users who would be negatively impacted. Push reduction: identifying the ~10% of users sensitive to fewer push messages prevents a drop in click‑through rates. Limited‑time coupons: a dual‑objective uplift model balances membership churn against playback time gains, shielding 20% of users with high negative impact.

PSM User Matching is used when A/B experiments are infeasible. The workflow includes problem definition, propensity‑score estimation and matching, balance checking, causal effect inference, and sensitivity testing. Several real‑world cases (operator‑sponsored membership, incentive ads, short‑term coupons, welfare center) demonstrate how PSM quantifies incremental value for activities lacking experimental traffic.

Spark CausalML addresses the limitation of single‑machine causal inference packages by providing a distributed PySpark implementation capable of handling millions of samples. It wraps popular causal inference methods for both experimental and observational data, has been open‑sourced internally at Tencent, and is being prepared for external release.

The presentation concludes with thanks to the speakers and organizers.