Applying Causal Inference Tools for Growth Scenarios in Industry

The presentation introduces causal inference tools tailored for growth scenarios in industry, emphasizing the need to move beyond simple correlation to true causation for reliable business decisions.

It explains why causal tools are required, using examples like movie releases versus drowning incidents and glasses versus academic performance to illustrate spurious correlations and the importance of distinguishing causal effects.

Data‑flow standards are covered, focusing on randomized controlled trials (RCTs), their pros and cons, nested versus non‑nested designs, the necessity of proper traffic shuffling, and the choice between user‑level and request‑level feature dimensions depending on the business case.

Model selection is discussed, highlighting common causal models, the principles of causal forests that maximize heterogeneity, generalized causal forests, policy‑learning approaches that tailor split criteria to ROI or Qini scores, and other de‑confounding techniques such as CBIV, feature decomposition, DML, and front‑door adjustment.

For model evaluation, metrics like AUUC and Gini curves are described, along with offline simulation pipelines that estimate uplift, cost‑benefit curves, and the integration of optimization under limited resources using dual‑problem formulations and greedy online allocation.

The end‑to‑end system workflow connects data collection, model training, online prediction, operations planning, and feedback loops, enabling continuous improvement of causal models and maximization of business revenue.