Description

The courselet Causal ML Primer introduces the foundational ideas and motivations behind causal inference in machine learning, emphasizing the crucial distinction between correlation and causation. Through real-world business examples, it illustrates how standard predictive models can mislead decisions when they fail to account for underlying causal structures. The lecture presents the potential outcomes framework as a formal basis for causal reasoning, outlines key causal parameters and assumptions, and discusses how to estimate treatment effects using both randomized and observational data. It further explores uplift modeling techniques for personalized decision-making and highlights methodological contrasts between econometrics and machine learning, paving the way for integrating both perspectives in causal ML applications.

1. Motivation: Correlation vs. Causation

• Traditional machine learning (ML) focuses on prediction, not causal inference.

• Spurious correlations often mislead ML models, especially when data reflect underlying decision processes (e.g., hotel pricing affected by demand).

• Real-world decision-making requires understanding cause-effect relationships, not just associations.

2. Illustrative Examples

• Hotel Booking Use Case: ML may learn that higher prices lead to more bookings—due to revenue management strategies, not a causal relationship.

• Churn Prediction: Targeting high-risk customers without considering past retention campaigns leads to collider bias—an artifact of conditioning on a variable affected by multiple causes.

3. Basics of Causal Inference

• Potential Outcome Framework:

  • Each unit has two potential outcomes: with and without treatment.

  • The Fundamental Problem: we only observe one of the two potential outcomes.

• Key Parameters:

  • ATE (Average Treatment Effect): average effect of treatment.

  • ATT (Average Treatment Effect on the Treated).

  • CATE (Conditional ATE): effect conditional on covariates.

• Key Assumptions:

  • SUTVA: no interference between units.

  • Ignorability: treatment assignment independent of outcomes, conditional on covariates.

  • Overlap: treatment and control groups must overlap in covariate space.

4. Study Designs

• Randomized Controlled Trials (RCTs): gold standard for causal inference, but costly and sometimes unethical.

• Observational Studies: more common but susceptible to bias from confounding, colliders, or selection.

5. Uplift Modeling

• Goal: estimate individual-level causal effects for decision-making (e.g., whom to target in a marketing campaign).

• Four types of customers:

  • Persuadables: respond only if treated.

  • Sure Things: respond regardless.

  • Lost Causes: never respond.

  • Do-not-disturbs: respond negatively to treatment.

6. Uplift Modeling Approaches

• Two-Model Approach: separate ML models for treated and control groups; uplift = difference in predicted outcomes.

• Interaction Model (Lo, 2002): single model with treatment-covariate interactions.

• Transformed Outcome Methods (Lai, etc.): encode uplift into target variable and train classifier.

• Uplift Trees: modify decision tree splitting to directly optimize uplift (e.g., via KL divergence).

7. Comparison: Econometrics vs. ML

Athey & Imbens (2019) highlight the cultural gap but also the opportunity for Causal Machine Learning to merge both perspectives.

Let me know if you’d like this summary in PowerPoint format or expanded into a lecture script.