Introduction to H2O Driverless AI by Marios Michailidis
- Predictive AI
- H2O Driverless AI
Overview
Marios Michailidis, a top-ranked Kaggle Grandmaster, introduces Driverless AI and its approach to automating feature engineering, model selection, and hyperparameter tuning. He walks through the standard data science workflow, where Driverless AI fits in once a target variable and tabular dataset exist, and how it balances accuracy against available time using an evolutionary, exploration-versus-exploitation search. The session covers feature transformations, feature importance, and the machine learning interpretability module used to explain complex model predictions.
Key takeaways
- How Driverless AI automates the iterative feature engineering, algorithm selection, and hyperparameter tuning steps of a data science workflow
- How to set an objective and balance accuracy against available time and hardware, with Driverless AI running an evolutionary search that explores new feature and algorithm combinations while exploiting what already works
- Common categorical feature transformations: frequency encoding, label encoding, one-hot encoding, and target encoding, with leave-one-out and Bayesian smoothing to avoid leakage
- How feature importance is measured by shuffling one feature at a time and observing the drop in validation accuracy
- How a final stacking step blends predictions from multiple algorithms, and how the machine learning interpretability (MLI) module uses simpler surrogate models to explain predictions with per-record reason codes
- A benchmark result: Driverless AI reached a top-10 finish out of 3,000 teams on a Kaggle competition in about two hours, versus roughly two to three weeks for a human competitor
Start owning your AI