Causal Inference in Wing Payment's Intelligent Decision-Making: Exploration and Practice

Introduction: The article explains what causal inference is, why it matters, and outlines the main topics covered, including definitions, challenges, and practical applications in intelligent decision‑making.

What is causal inference: It is the scientific study of identifying causal relationships between variables, distinguishing causation from mere correlation, illustrated with examples of spurious correlations (e.g., chocolate consumption vs. Nobel prizes, butter consumption vs. divorce rates).

Key challenges: The analysis must address confounding variables and selection bias, which can create false correlations; these factors are explained with classic scenarios such as Simpson's paradox and clinical treatment comparisons.

Main applications and methods: The article describes two major problem types—causal discovery (building causal graphs) and causal effect estimation (ITE, ATE, CATE). It introduces common estimation techniques, including matching, difference‑in‑differences, synthetic control, and uplift modeling.

Frameworks: Two dominant frameworks are presented—the Potential Outcome Framework (Rubin) for effect estimation and the Structural Causal Model (SCM) for constructing causal graphs.

Practical scenarios in Wing Payment: (1) Evaluating the impact of a new feature (e.g., homepage pop‑up) using double‑difference methods when A/B testing is costly; (2) Building a response model to predict conversion and a uplift model (S‑learner meta‑learner) to estimate individual treatment effects; (3) Applying class‑transformation methods for binary outcomes. Model evaluation metrics such as AUUC are discussed.

Personal reflections: The author recommends using A/B tests whenever feasible, notes that uplift models may not always outperform response models, and emphasizes the importance of proper experimental design and variable independence.

Q&A: Answers cover how to set thresholds for response models, handling missing treatment variables in S‑learner (suggesting T‑learner), and evaluation metrics for uplift models (AUUC, QINI).