Repository navigation
v0.3.11
Add configurable neighbor weights for both complete and available donor policies.
- Support
weights="uniform"(default),"distance", and callable weights.Noneis also accepted as uniform weighting. - Apply inverse Euclidean distance weighting, with selected zero-distance donors taking precedence for each missing feature.
- Preserve existing uniform mean/median aggregation and unweighted fallback statistics.
- Normalize weights and use float64 weighted aggregation to reduce overflow risk.
- Validate callable output shapes, weight values, and zero weight sums.
Distance and callable weights require strategy="mean" and metric="l2".
Validation: 303 tests passed, including 59 new weighted-imputation cases. Two warnings were reported by existing complete-aggregation overflow tests.