Preprocessing 0.2.0 knnimputeroptions

Lodestar.Preprocessing 0.2.0. This page is frozen at that release. Read the current documentation for what main says now. A link to a decision or a migration page follows main, and leaves the archive.

Home › Encoding and imputation

KnnImputerOptions

How many neighbours KnnImputer averages, and how it weights them.

public sealed record KnnImputerOptions

Properties — NeighbourCount is how many donors each missing value is averaged over; scikit-learn's n_neighbors, default 5. Weights is whether nearer donors count for more; weights, default NeighbourWeights.Uniform.

Example — two donors rather than five, on a four-row matrix.

using Lodestar.Preprocessing;

double[] rows =
[
    1.0, 2.0, double.NaN,
    3.0, 4.0, 3.0,
    double.NaN, 6.0, 5.0,
    8.0, 8.0, 7.0,
];

var options = new KnnImputerOptions { NeighbourCount = 2 };

double filled = KnnImputer.Fit(rows, 3, options).Transform(rows)[2];   // => 4

Remarks — the neighbour count is a ceiling, not a requirement. Asking for five donors on a matrix where only three rows carry the feature averages those three; nothing is refused and no row is padded.

metric is absent. The reference publishes one value for it, 'nan_euclidean', and documents that it is the only one supported — so there is no choice to carry. missing_values is absent with it: NaN is the marker, which is the reference's own default and the only one that needs no sentinel value hunting through a column of doubles. copy and add_indicator are absent for the reasons the rest of this package gives — a span goes in and a new array comes out, and an indicator column is something a caller composes.

keep_empty_features is absent because the default is the only defensible answer here: a feature missing from every fitted row is dropped, as KnnImputer.Fit describes.

Being a record, two option sets with the same two values are equal.

Applies to — net10.0, netstandard2.0.

See also — KnnImputer.Fit, NeighbourWeights.