fix: count/nunique aggregates yield 0, not NaN, for parents without children

[CristobalSantana]

What

In a one_to_many relation, aggregates are computed by grouping the child table and left-merging onto the parent. Parents with no child rows therefore receive NaN in every aggregate column, including count and nunique. Downstream, preprocessing bins those NaNs as missing.

The count of an empty set is 0, not a missing value. Leaving it as NaN conflates "no children exist" (a real, often highly informative value) with "data is missing", and the two become indistinguishable in the trained network. This felt like exactly the kind of hidden behavior the design principles ask contributors to avoid.

How I found it

I stress-tested the library with a synthetic public-tenders dataset (3,000 tenders, 8,000 bids, target = tender declared void). 287 tenders had zero bids, which was by construction the strongest predictor of the target (80% positive rate vs 22% global). After materialization, all 287 had bid_count = NaN and were binned as missing. With the fix, on that dataset, ROC-AUC went from 0.753 to 0.808 and PR-AUC from 0.533 to 0.608. The metric gain is dataset-specific, but the semantic issue is general: in real data that also has genuinely missing child records, the two cases cannot be separated at all without this distinction.

Change

After the aggregate merge, fill NaN with 0 only for columns produced by count and nunique ops. Other ops (mean, latest, etc.) still yield NaN, since they are truly undefined over an empty set. This also covers windowed counts (window_days), where a parent with children outside the window now gets 0 instead of NaN.

Note this is a behavior change for anyone relying on NaN counts; given 0.1.0 alpha status I assumed the semantic fix is preferred, but happy to gate it behind a config flag if you'd rather keep the current default.

Testing

New regression test: parent with zero children gets count == 0, nunique == 0, and mean stays NaN. uv run ruff check . clean, full test suite passes (63 tests), smoke test auto-bayesian train examples/lead_scoring.toml runs.

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[CristobalSantana]

I have read the CLA Document and I hereby sign the CLA