Lodestar.Cluster 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 › Partitioning

Clusters from a square distance matrix rather than from the samples themselves.

public static Dbscan FitPrecomputed(ReadOnlySpan<double> distances, int sampleCount, double epsilon, int minimumSamples)

Parameters — distances is the pairwise distances, row-major and square. sampleCount is the side of that matrix. epsilon is the inclusive radius of a neighbourhood, scikit-learn's eps. minimumSamples is how many samples a neighbourhood needs to be dense, the sample itself counted.

Returns — a fitted Dbscan.

Exceptions — ArgumentOutOfRangeException when sampleCount or minimumSamples is not positive, or epsilon is not positive or not finite. ArgumentException when distances is not sampleCount squared values, or holds a NaN or infinite one — an infinite distance is refused too, as the reference refuses it, rather than read as unreachable.

Example — three samples, given as the distances between them.

using Lodestar.Cluster;

// Row-major and square: the first two samples are one apart, the third is far from both.
double[] distances =
[
    0.0, 1.0, 5.0,
    1.0, 0.0, 4.0,
    5.0, 4.0, 0.0,
];

Dbscan model = Dbscan.FitPrecomputed(distances, sampleCount: 3, epsilon: 1.0, minimumSamples: 2);

int pair = model.Labels[0];      // => 0
int alone = model.Labels[2];     // => -1
int clusters = model.ClusterCount;   // => 1

Remarks — this is metric="precomputed", and it answers exactly what Dbscan.Fit answers on the samples those distances came from. The corpus freezes one case both ways to hold that true rather than to assert it.

The diagonal is a zero distance, so a sample is its own neighbour here too — which is what keeps one meaning of minimumSamples across both entry points. A matrix whose diagonal is not zero is not rejected, because nothing here requires a metric: it is read as given, which is also what lets a caller pass a dissimilarity that is not a distance at all.

Named rather than an overload of Fit. Both take a row-major ReadOnlySpan<double> and an int, so an overload pair would be told apart only by what the second argument means — and a caller who passed a sample matrix where a distance matrix was expected would get an answer rather than an error.

Applies to — net10.0, netstandard2.0.

See also — Dbscan, Dbscan.Fit, clustering metrics.