Skip to content

v0.3.0

Choose a tag to compare

@ScionKim ScionKim released this 03 Sep 08:22
· 118 commits to main since this release
ba72540

New features

  • Add opt-in donor_policy="available" support for partially observed training donors.
  • Select donors separately for each missing feature, using shared observed features for distance calculation.
  • Use fitted column statistics when no eligible donor is available.
  • Keep donor_policy="complete" as the default, preserving existing behavior.

The available-donor policy supports metric="l2", index_factory="Flat", and both mean and median aggregation.

Reliability

  • Add batched NaN-aware distance calculations and numerical safeguards.
  • Expand regression tests, dependency compatibility checks, and installed-package smoke tests.

Performance

Performance depends on the data and missingness pattern; this mode is not guaranteed to outperform KNNImputer.

The partial-donor benchmark report measures development commit 0a3cc077, not the final 0.3.0 implementation.