It's the Voronoi algo more or less verbatim which I initially read as an alternative implementation of PAM as opposed to an alternative to PAM. I'll add the following bits to it and then put it up on Github if it piques an interest:
1. I'll treat the distance calculation as an adverb to avoid having to instantiate the whole distance matrix (which is never used in its entirety). This will give it better space complexity. I'll try to make it so that it could be run on say 1M data points. 2. I'll add kmeans++ initialisation which should perform much better than random. 3. I'll use a Gaussian mixture model with spherical variance parameterised by the medoids and sample variance in-cluster as a working model to assign likelihood to fitted data. 4. I'll add repeated initialisations. I.e. repeat clustering N times and keep the initialisation with the best likelihood. 5. I'll add a k discovery routine which uses AIC and the GMM likelihood to find k. 6. I'll include benchmarks against kmeans, kmedoids, affinity prop and a couple other sklearn implementation against a couple datasets. ---------------------------------------------------------------------- For information about J forums see http://www.jsoftware.com/forums.htm
