Development build. This page describes main, not a released package. The latest published Lodestar.Preprocessing is 0.2.0 — read its documentation.

Home › Preprocessing › Encoding and imputation

KnnImputer

Fills each missing value from the rows most like the one it is missing from, at sklearn.impute.KNNImputer parity.

public sealed class KnnImputer

Properties — FeatureCount and SampleCount are the shape it was fitted on. OutputFeatureCount is how many values a filled row carries, and KeptFeatures says which feature each one is — fewer than FeatureCount when a feature was missing from every fitted row, since there is then nothing to impute it from and nothing to impute it with.

Example — a NaN in each of two rows, filled from the two nearest donors.

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,
];

KnnImputer imputer = KnnImputer.Fit(
    rows, featureCount: 3, new KnnImputerOptions { NeighbourCount = 2 });

double[] filled = imputer.Transform(rows);

double firstGap = filled[2];   // => 4
double secondGap = filled[6];  // => 5.5

Remarks — this is the imputer that reads the row, where SimpleImputer reads the column. A mean fills every gap in a feature with the same number; this one asks which other rows resemble the one with the gap and answers with theirs, so a row that looks like the small ones gets a small value and a row that looks like the large ones gets a large one. Above, the missing third feature of the first row becomes 4 — the mean of its two nearest donors — not 5, the column mean.

Distances are nan_euclidean. The squared differences are taken over the coordinates both rows have and scaled up by the fraction that were missing, so a pair sharing two features of four is not automatically nearer than a pair sharing all four. It is the reference's own metric.

It costs a distance between every pair, so the work is rows × fitted rows × features — which is why the call is refused past a measured ceiling rather than run for however long that takes. Transform has the number.

Applies to — net10.0, netstandard2.0.

See also — KnnImputerOptions, NeighbourWeights, SimpleImputer, the encoding index, the Python equivalence table.

Members

Member	What it does
KnnImputer.Fit	Keeps the fitted rows, which are the donors.
KnnImputer.Transform	Fills every missing value from its nearest donors.