Repository navigation
v0.3.1
Performance
- Optimize the complete-donor data layout to reduce repeated column-projection overhead.
- Add reproducible missingness-pattern benchmarks, raw results, and a report.
- Keep the public API, imputation algorithm, and available-donor policy unchanged.
Performance gains depend on the workload. Small training sets with many distinct missingness patterns can still be slower than KNNImputer.
See the [complete-donor benchmark report](https://github.com/ScionKim/FaissImputer/blob/v0.3.1/docs/benchmarks/complete-patterns-9d179b2b.md) for measurements and limitations.