The ZIP Code That Ate the Cluster

Adding RAM treats a symptom. H2O holds entire frames in memory, column-compressed, so a 9.2 GB CSV typically expands to roughly 2–4x that once parsed, and categorical domains add per-level metadata. Worse, several algorithms expand categoricals into a design matrix internally - H2O's XGBoost wrapper one-hot encodes categorical columns, and GLM does too - so zip_code alone can become ~41,000 columns and pro_id ~310,000 over 11.8M rows. Cross-validation then multiplies the footprint: each fold holds training frames plus a model copy, and retaining CV predictions for stacked ensembles keeps more alive at once. That is the OOM.

Prescriptions:

• pro_id: drop it. It is an identifier with hundreds of thousands of levels, near-unique relative to the signal it carries; it will memorise professionals rather than generalise, and it is useless for a quote that has not yet been matched.
• zip_code: recode. Roll up to 3-digit prefix or a sales region (hundreds of levels), or replace with frequency and a small set of market-level aggregates. Keep the raw code only as an ignored_columns entry.
• device_model: collapse. Top-N models plus an 'other' bucket. Then set ignored_columns explicitly so the raw columns never enter AutoML, reduce nfolds while iterating, and disable keep_cross_validation_predictions until the final run.

days_since_signup, 31% missing. Missingness this large is usually a signal, not noise - accounts created in-session have no value. Do not mean-impute: it collapses a meaningful group into the middle of the distribution. Add an explicit signup_missing binary flag, and let GBM/DRF/XGBoost route NAs down a default branch, which they handle natively. If the downstream model cannot take NAs (e.g. a plain GLM), impute a sentinel and keep the flag.

quote_amount. Tree models are invariant to monotone transforms, so the skew is harmless for GBM/DRF/XGBoost - log-transforming buys nothing and costs interpretability. The negatives are the real issue: 6% refunds are a different business event, not a small quote. Create is_refund and model the absolute amount separately, or clip and flag. If a regularised GLM is in the mix, rank-transform or winsorise, because IRLSM will be dominated by the tail.

quote_timestamp. Force parsing at import (col.types with a date type, or h2o.as_date with the exact format) rather than leaving it as a string; a string column is either dropped or one-hot encoded into hundreds of thousands of garbage levels. Then derive hour-of-day, day-of-week, month, holiday proximity, and days-since-previous-quote. Do not feed the raw epoch as a numeric feature - trees cannot extrapolate beyond the training range and the model will silently go stale.

The decision hinges on who owns the encoding work and how much transparency is required.

H2O-3 gives explicit control: you decide exactly how zip_code is recoded, what is dropped, and which folds are used. That is valuable here because the correct treatment (drop the professional ID, roll up geography) is a domain judgement, not a statistical one, and you want it documented and reproducible. The cost is that AutoML will happily one-hot encode whatever you leave in the frame.

Driverless AI automates much of this: it detects high-cardinality categoricals and applies frequency and out-of-fold target encoding rather than one-hot expansion, adds date and text feature engineering, and runs time-boxed experiments with its own model and hyperparameter search. For a dataset with three very high-cardinality fields, DAI is likely to reach a usable model faster without manual recoding - but the target encoding is a statistical transform that must be audited for leakage and for stability as new levels appear, and the resulting pipeline is less transparent to a sceptical reviewer.

Either way, the non-negotiable step is the same: drop pro_id before training. Neither tool should be trusted to discover that on its own. If the team's priority is explainability and a documented feature pipeline, choose H2O-3; if it is time-to-first-model and the team lacks feature engineering capacity, choose DAI and budget time for encoding review.