CKMeans ported from C++ to C#, originally used in R language.
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).
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.
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).
xare points / positionsyare weightsKminis minimal number of clustersKmaxis maximal number of clustersmethodis what method to use for calculation (Linear,LogLinear,Quadratic).LogLinearis fastest.
All functions return CKResult object, which contains:
int[] Clustersis an array of cluster IDs for each point in x,number[] Centersis an array of centers for each clusternumber[] Withinssis an array of within-cluster sum of squares for each clusternumber[] Sizesis an array of (weighted) sizes of each clusterdouble[] BICis Bayesian Information Criterion
Tutorial "Optimal univariate clustering" (for R)
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.