Sharp-CKMeans

CKMeans ported from C++ to C#, originally used in R language.

What is this?

This is library for 1-dimensional/univariate optimal clustering.

It's special because it's deterministic - it doesn't use random numbers & iteration for calculation, which means it always returns same result for same input, it's precise (always most optimal result) and very fast (uses dynamic programming).

Why use this?

Works natively in C#. Made so that it works with decimal data type.

You can use this instead of Gaussian Mixture Model (depends on what you wish to do though).

It's slow with decimal though (takes around 3ms per clustering).

As an alternative, you can change using number = System.Decimal; to using number = System.Double; on top of every file, this way it will work with doubles (or floats if you wish) and will be faster, although less precise.

How to use

All needed functions are present at Main.cs.

What you have available is Main.CKMeans1D(...), Main.CKMedian1D(...), Main.CKSegs1D(...).

For all functions, arguments are (number[] x, number[] y, int Kmin, int Kmax, Method method = Method.Linear).

x are points / positions
y are weights
Kmin is minimal number of clusters
Kmax is maximal number of clusters
method is what method to use for calculation (Linear, LogLinear, Quadratic). LogLinear is fastest.

All functions return CKResult object, which contains:

int[] Clusters is an array of cluster IDs for each point in x,
number[] Centers is an array of centers for each cluster
number[] Withinss is an array of within-cluster sum of squares for each cluster
number[] Sizes is an array of (weighted) sizes of each cluster
double[] BIC is Bayesian Information Criterion

Tutorial "Optimal univariate clustering" (for R)

Log-Likelihood

Logarithm/Exponent (log likelihood) calculation is done with doubles (64-bit in C#), originaly it was done with long-double (80-bit in C++).

I added tests for CKMeans (not CKMedian and others!) and it works as expected, but if you experience irregularities, you need to convert log-likelihood calculation part of code to higher precision. I tried with decimal but it threw out of range errors!

It's short part of code, it's present in NonWeighted.cs at line 150 and Weighted.cs at line 145.

Original sources

Original page for this library

Original Github mirror

viliwonka / Sharp-CKMeans