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

Home › Preprocessing › Feature transforming

Fits bin edges on a row-major matrix.

public static KBinsDiscretizer Fit(ReadOnlySpan<double> samples, int featureCount, KBinsDiscretizerOptions options = null)

Parameters — samples is the matrix, row-major, featureCount values per row; the span is read, never modified. options chooses the bin count, the strategy, the encoding and the percentile convention; null takes the reference's defaults, which are five quantile bins encoded one-hot.

Returns — a fitted KBinsDiscretizer.

Exceptions — ArgumentOutOfRangeException when featureCount is not positive or the bin count is below two. ArgumentException when samples holds no row, a partial one, or a non-finite value.

Example — the same column under two strategies, which do not agree.

using Lodestar.Preprocessing;

double[] ages = [19.0, 22.0, 25.0, 31.0, 38.0, 44.0, 52.0, 61.0, 67.0, 74.0];

KBinsDiscretizer byCount = KBinsDiscretizer.Fit(
    ages, 1, new KBinsDiscretizerOptions { BinCount = 4, Encoding = BinEncoding.Ordinal });

KBinsDiscretizer byWidth = KBinsDiscretizer.Fit(
    ages, 1, new KBinsDiscretizerOptions
    {
        BinCount = 4,
        Strategy = BinStrategy.Uniform,
        Encoding = BinEncoding.Ordinal,
    });

string equalCounts = string.Join(",", byCount.Transform(ages));  // => 0,0,1,1,1,2,2,3,3,3
string equalWidths = string.Join(",", byWidth.Transform(ages));  // => 0,0,0,0,1,1,2,3,3,3

Remarks — the strategy is the decision, not the bin count. BinStrategy.Quantile gives every bin the same number of rows and lets the widths vary; Uniform gives every bin the same width and lets the counts vary, which on a skewed column puts most of the data in the first bin; KMeans lets the data decide both, at the cost of an iterative fit.

The edges are numpy's to the last bit. Uniform edges are np.linspace(min, max, n + 1) and the quantile levels np.linspace(0, 100, n + 1), each percentile read through numpy's own _lerp: an edge an ulp off the reference's moves a value sitting on it into the next bin (#1226, #1237). A constant feature is one bin over [-inf, inf], as the reference makes it, so its inverse is NaN (#1230).

The fit reads every row of every feature and sorts each column, so it costs features × n log n. There is no subsample here, for the reason QuantileTransformer has none either.

Applies to — net10.0, netstandard2.0.

See also — KBinsDiscretizer.Transform, KBinsDiscretizerOptions, the Python equivalence table.