Causal Machine Learning for User Growth: Concepts, Methods, and Applications

In this talk, guest speaker Li Ao (PhD, Kuaishou) discusses the integration of causal reasoning and machine learning to address subtle correlations in massive data and evaluate predictive accuracy for user growth.

1. Basic Concepts – User growth metrics focus on DAU, retention, activity, and market-driven indicators such as payment and viral spread.

2. Causal Analysis Methods – Propensity Score Matching (PSM) is introduced to answer "why" questions, while causal machine learning and attribution techniques (Uplift, Meta‑learner, Causal Recommendation) address the "how".

3. PSM Procedure – (1) Build a treatment model (e.g., LR/XGBT) to estimate propensity scores; (2) Match treatment and control groups to remove bias; (3) Perform KS test to check covariate balance; (4) Compute Average Treatment Effect (ATE) to quantify impact, e.g., a 5% lift in retention from clicks.

4. Causal Machine Learning – Differentiates between using causal inference as a tool for ML and using ML for causal inference; discusses heterogeneous treatment effect (HTE) analysis for retention and activity, illustrating quadrant classification of treatment effects.

5. Game‑Coin Recovery Model – Presents a mathematical model comparing two recovery strategies (100 coins vs. 60 coins) and evaluates them using Meta‑learner and HTE approaches, highlighting the calculation of lift and optimal policy selection.

6. Causal Attribution Theory – Addresses challenges of multi‑treatment scenarios and delayed effects, proposing a framework that uses propensity scores and credit assignment to attribute lift to specific treatments.

7. Q&A Highlights – The causal approach is applied mainly in the re‑ranking stage of recommendation pipelines, can be extended to sorting, and benefits from random traffic for unbiased estimation.

Overall, the talk demonstrates how causal inference techniques can be systematically integrated into user growth strategies to improve retention, activity, and monetization.