Causal machine learning for a smart paywall at The New York Times
summary
Claims a deployed prescriptive system called Dynamic Meter that uses causal meta‑learners trained on randomized control trials to pick per‑user meter limits, trading off subscription conversion versus engagement with a tunable delta. Evaluation uses inverse‑probability weighting to estimate counterfactuals and shows a higher conversion rate at the same engagement (a Pareto improvement). Production stack runs Python models, schedules with Airflow and Docker, and reads/writes to Google BigQuery; the model is live and yields measurable lift.
AI-written summary · about
chapters (10)
- 0:00Causal ML at NYT
- 3:11Data science at NYT
- 7:27NYT paywall evolution
- 10:56Dynamic meter project
- 14:26Predictive vs prescriptive ML
- 16:12Dynamic meter as causal model
- 17:36Engagement and subscriptions modeling
- 21:25Model evaluation with retrospective analysis
- 25:16NYT ML tech stack
- 27:23Q&A on fairness
related talks
38mDeploying Data Science for Distribution of The New York Times
38mData science at The New York Times
43mBeyond Just Prediction: Causal Thinking in Machine Learning
06mCausal Machine Learning – What’s in it for Data Scientists?
39mMachine Learning Engineering
23mThe unreasonable effectiveness of data for better monetisation
33mMachine Learning to moderate ads
30mReconceptualizing Machine Learning for the Real-time World