Intro to AutoML + Hands-on Lab - Erin LeDell, Machine Learning Scientist, H2O.ai
- Predictive AI
- H2O Driverless AI
- H2O-3
Overview
In this talk from H2O World 2017, machine learning scientist Erin LeDell introduces H2O's approach to automatic machine learning, combining random grid search with stacked ensembles. She explains how AutoML handles data preprocessing, hyperparameter tuning, and early stopping, and how the resulting leaderboard ranks models, including an all-models ensemble and a leaner best-of-family ensemble. The session then moves into a hands-on lab running AutoML in Python and R on binary classification and regression datasets.
Key takeaways
- What automatic machine learning covers across the pipeline: data preparation, model generation through grid or random search, and combining models into ensembles
- Why H2O AutoML uses random grid search over Cartesian grid search, plus early stopping on individual models and on the grid itself, to find good models faster
- How stacked ensembles work: cross-validated predictions from base learners become training data for a meta learner, producing both an all-models ensemble and a smaller best-of-family ensemble for production use
- The algorithms covered in H2O AutoML at the time, GBM, random forest, deep neural nets, and GLM, with XGBoost noted as coming soon, and the resulting leaderboard ranked by cross-validated performance
- A live hands-on demo training AutoML on a product backorder binary classification dataset and a power plant regression dataset, including how to inspect meta learner variable importance to see which base models contribute most
- The H2O-3 roadmap items mentioned at the time, including Cox proportional hazards, ordinal regression, custom meta learners, and planned GPU support
Start owning your AI