Preprocessing 0.2.0 neighbourweights

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

NeighbourWeights

How KnnImputer weights the donors it averages.

public enum NeighbourWeights

Values — Uniform counts every donor the same; scikit-learn's 'uniform', and the default. Distance counts each donor as the reciprocal of its distance; 'distance'.

Example — the same gap, filled both ways.

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 uniform = new KnnImputerOptions { NeighbourCount = 2 };

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

KnnImputer weighted = KnnImputer.Fit(
    rows, 3, uniform with { Weights = NeighbourWeights.Distance });

double nearer = Math.Round(weighted.Transform(rows)[2], 4);          // => 3.6667

Remarks — Distance matters when the donors are not equally close, which is most of the time: above, the nearer donor carries a 3 and the further one a 5, so weighting pulls the filled value from 4 down to 3.67. Prefer it when the rows are genuinely comparable on a metric scale, and the default when they are not — an unweighted mean of the nearest few is the more robust answer on a matrix whose features were not scaled.

A donor at distance zero takes the whole weight. A row identical on every shared feature gives an infinite reciprocal, and the reference resolves that by averaging only the exact matches and ignoring the rest; so does this.

Applies to — net10.0, netstandard2.0.

See also — KnnImputerOptions, KnnImputer.Transform.