Feature Engineering with H2O - Dmitry Larko, Senior Data Scientist, H2O.ai
- Predictive AI
- H2O Feature Store
Overview
H2O senior data scientist and Kaggle competitor Dmitry Larko gives a deep, practical tour of feature engineering: what counts as feature engineering versus data cleaning, and the techniques that separate a good tabular model from a great one. He covers categorical encoding methods from one-hot to target mean encoding with smoothing, numerical transformations, feature extraction from GPS and datetime fields, and answers extensive audience questions on validation, missing values, and time series pitfalls.
Key takeaways
- The difference between feature engineering and data cleaning or scaling, and why target and business context should drive most feature choices
- Categorical encoding options and their tradeoffs: one-hot, label, frequency, target mean encoding with smoothing for rare categories, leave-one-out, and weight of evidence
- How to use information value as a rule of thumb for judging a feature's predictive power, and as a warning sign for data leakage
- Numerical feature techniques: quantile binning, dimensionality reduction such as PCA or SVD, clustering, and distance-to-centroid features
- How to extract new features from raw fields, such as mapping GPS coordinates to zip codes or pulling year, month, day, and holiday flags from timestamps
- Why validation discipline matters most for time series features, so training features never leak information from the prediction window
Start owning your AI