Automatic Feature Engineering with Driverless AI
- Predictive AI
- H2O Driverless AI
- H2O Feature Store
Overview
Dmitry Larko, a top-100 ranked Kaggle Grandmaster, explains how Driverless AI automates feature engineering, the step he argues most separates good data scientists from the best. He covers why the right features let simpler models solve problems that would otherwise need much more complex ones, plus practical techniques for encoding categorical, numerical, and text features. A live demo runs Driverless AI against a Two Sigma rental dataset, followed by audience Q&A on overfitting and imbalanced data.
Key takeaways
- Why feature engineering, illustrated with a simple example of reshaping data from Cartesian to polar coordinates, can let a simpler model solve a problem that would otherwise need a much more complex one
- Categorical encoding techniques: label encoding, one-hot encoding, frequency encoding, and target encoding, with leave-one-out and Bayesian smoothing to avoid leakage on high-cardinality features
- Numerical and text feature techniques: log-transforming the target variable, binning into quantiles or histograms, and encoding text via TF-IDF and word2vec-style embeddings
- How Driverless AI finds feature interactions through a mix of random search and feedback from gradient boosted models, balancing exploration against exploiting what already works
- How a "reliable holdout" technique guards against overfitting by only keeping a new transformation if it produces a statistically significant score improvement
- A live demo on a Two Sigma rental listings dataset, where price, location, and building or manager ID (via target encoding) emerge as the strongest predictors of listing interest
Start owning your AI