Scalable Automatic Machine Learning in H2O
- Predictive AI
- H2O Driverless AI
Overview
In this webinar, H2O.ai chief machine learning scientist Erin LeDell explains what automatic machine learning is and how H2O's open-source H2O AutoML tool implements it. She covers data preprocessing, the random grid search approach H2O uses across GBM, random forest, extremely randomized trees, GLM, and deep learning models, and the two stacked ensembles AutoML produces (all models and best of family). The session includes practical tuning tips (validation frames, max_models versus max_runtime_secs, memory limits, early stopping, project_name) and a live tutorial in Flow, R, and Python using a product backorders dataset.
Key takeaways
- What automatic machine learning covers: data preprocessing, generating many models through grid search, and ensembling them together
- Why H2O AutoML favors a large random grid search across GBM, random forest, extremely randomized trees, GLM, and deep neural nets, and how that diversity feeds its stacked ensembles
- The difference between the two ensembles AutoML returns: an all-models stacked ensemble (best performance) and a best of family ensemble (five models, close performance, faster to score)
- Practical run settings: how to use a validation frame instead of a leaderboard frame to avoid wasting data, when to use max_models with a seed for reproducibility versus max_runtime_secs, and how to add more models to an existing run using project_name
- How to run AutoML in H2O's Flow GUI, R, and Python against a product backorders dataset, and how to read the resulting leaderboard, ranked by cross-validated AUC and log loss
- Where AutoML differs from Driverless AI, and what's on the roadmap, including XGBoost integration and target encoding
Start owning your AI