Notes
🔍

H2O Driverless AI

If H2O AutoML is the open-source engine, H2O Driverless AI is the commercial vehicle built around it: a licensed platform you run on your own GPU server or cloud tenancy, aimed squarely at the parts of the job that model search alone does not cover — features, explanations and deployment.

Feature engineering as the main event

Its distinguishing machinery is an evolutionary loop over feature transformations rather than only over models. A genetic algorithm proposes, recombines and discards transformers, scoring each candidate feature set by the accuracy it actually buys:

  • target and frequency encodings for high-cardinality categories;
  • cluster distances and truncated SVD over groups of numeric columns;
  • lags, rolling windows and date expansions for time series;
  • text vectorisers and pretrained embeddings for free-text columns.

Three dials and a shippable artefact

Instead of algorithm-level knobs you set accuracy, time and interpretability, the last of which restricts the search to features and models a human can still read. A finished experiment produces an Autoreport documenting what was tried and a MOJO scoring pipeline — a standalone artefact carrying the fitted transformations, deployable in Java or C++ with no Python runtime, which is the real reason regulated enterprises buy it. Built-in interpretability adds Shapley values, surrogate decision trees and partial-dependence plots. The trade-offs are the obvious ones: proprietary and licence-priced, GPU-hungry, and automatically generated features that can be awkward to defend one by one even when the report explains how they were made.

H2O Driverless AI is H2O.ai's commercial AutoML platform, licensed to run on your own server or cloud tenancy. It automates feature engineering through an evolutionary search over transformers as well as model selection, governed by three dials — accuracy, time and interpretability — and exports a MOJO scoring pipeline that runs without Python. It matters where enterprises need automated modelling with an audit trail: each experiment yields documentation and built-in interpretability for the resulting model.

Most of the skill in predicting something from a spreadsheet is not the predicting. It is working out which columns to combine, split or rescale so the pattern becomes visible at all.

H2O Driverless AI is a paid business product that automates that part. It tries thousands of ways of reworking your columns, keeps the ones that help, chooses a prediction method, and then writes a report explaining what it did — the kind of paperwork a bank or hospital needs before anyone is allowed to use the model. It also packages the finished model so the company's existing software can run it. You set three sliders: how accurate, how fast, and how easy to explain.