Applying Causal Inference to Business Improvement: Concepts, Methods, and Case Studies from Xiaohongshu

Introduction – The talk outlines the necessity of causal inference in business, defines the concept, and presents three main sections: why it is needed, what it is, and how it drives business improvement.

Why Causal Inference Is Needed – Traditional insight‑driven analysis provides correlations but cannot explain causality; causal inference offers a scientific framework (empirical evidence and logical reasoning) that enables reproducible, quantitative decision‑making, illustrated with the example of improving new‑user retention on Xiaohongshu.

What Causal Inference Is – Three academic streams are described:

Computer Science: Judea Pearl’s causal graph model (chain, fork, collider) and the do‑operator, with examples of back‑door adjustment in A/B testing.

Econometrics: Definition of treatment and counterfactual outcomes, selection bias, and methods such as Double Machine Learning (DML) to remove bias.

Statistics: Rubin’s Potential Outcome Model, average treatment effect (ATE), individual treatment effect (ITE), and how they combine with machine‑learning predictions.

How Causal Inference Drives Business Improvement

1. Improving new‑user retention – By modeling user behavior over time and attributing changes to specific interventions (e.g., internal video playback), causal methods such as Delta‑DAU or inverse‑probability weighting generate contribution matrices that guide content and product strategies, leading to measurable gains in retention metrics.

2. Increasing content creation – Estimating the causal effect of a user’s click on subsequent posting involves predicting a prior click‑to‑upload rate, attributing influence to specific notes, and identifying high‑impact content for recommendation, thereby encouraging more user‑generated posts.

The discussion also covers credit assignment, heterogeneous treatment effect estimation, and practical advice on selecting appropriate causal models for complex, real‑world data.

Conclusion – Causal inference provides a rigorous yet practical toolkit for turning insights into actionable business value, though models must be continuously validated against evolving data and operational constraints.