Causal AI and Uplift Modeling in H2O
- Predictive AI
- H2O Driverless AI
- H2O-3
This lesson demonstrates uplift modeling for causal AI, which measures the difference in response between a treatment group and a control group, such as customers targeted or not by a marketing campaign. It shows how to run uplift models in Driverless AI using custom recipes, including LightGBM and XGBoost uplift variants, a GLM, and a custom area under the uplift curve scorer, with a stratification check confirming the groups are balanced. It then walks through the same workflow natively in H2O-3 using the Uplift Random Forest, reviewing variable importance and predicted uplift.
Key takeaways
- How uplift modeling measures the difference in response between a treatment group and a control group
- What custom recipes are available for causal AI in Driverless AI, including LightGBM and XGBoost uplift variants and a GLM
- Why a stratification check confirms that treatment and control groups are randomly selected and comparably sized
- How H2O-3 offers a native distributed Uplift Random Forest through its Python library or the H2O Flow interface
- What the resulting output includes, predicted response with and without treatment and the calculated uplift per record
Start owning your AI