Beyond Just Prediction: Causal Thinking in Machine Learning
summary
If prediction and randomized AB tests disagree, causal methods estimate the effect of actions not just outcomes. Uplift modeling uses observed numbers like 20% purchase with email versus 8% without to flag a naive 12 point lift then shows confounding by engagement (high engagement: 20% vs 16% → 4 points) and low engagement (6% vs 3% → 3 points). The pipeline shows uplift modeling, practical estimators as s-learner and t-learner, evaluation by uplift-by-bin, and tooling examples like causalml and econML on a Cryio ad trial.
AI-written summary · about
chapters (16)
- 0:00Causal thinking in ML
- 2:03Correlation vs causation in ML
- 4:53Predictive vs causal models
- 8:45Randomized control trials gold standard
- 10:11Uplift modeling introduction
- 10:54Potential outcomes framework
- 16:12Customer response types
- 17:17Conditional average treatment effect
- 20:08Metalearners for treatment effects
- 21:13S-learner method
- 24:20T-learner method
- 27:11Evaluating uplift models
- 31:44Deploying uplift models
- 33:29Criteo dataset example
- 34:36Key takeaways
- 37:04Q&A
related talks
06mCausal Machine Learning – What’s in it for Data Scientists?
38mHow to Enhance Customer Targeting in Marketing
32mWhy start using uplift models for more efficient marketing campaigns
39mCausal Effect Estimation in Practice: Lessons Learned from E-commerce & Banking
44mIntroduction to Uplift Modeling
34mMachine Learning Who to Nudge: Causal vs Predictive Targeting in a Field Experiment on Student...
32mCausal Forecasting: How to disentangle causal effects, while controlling for ...
11mCausal Inference For Socio-Economic & Engineering Systems