ANNImputer#

ANNImputer builds an approximate vector index and retrieves candidate neighbors to fill missing features. The implementation supports Annoy and Voyager-oriented paths and exposes controls such as n_neighbors, metric, n_trees, search_k, weighting, initial strategy, and on-disk index build.

Approximation and preprocessing#

Rows with missing values still need a query representation suitable for the selected vector backend. Initial imputation/normalization choices therefore influence which neighbors are retrieved. Treat those choices as part of the model pipeline and evaluate them using held-out data.

Do not compare ANN and exact KNN results by expecting every imputed value to be identical. Compare downstream quality, latency, memory, and reproducibility under a fixed configuration.