ISO-SPLIT is an efficient clustering algorithm that handles an unknown number of unimodal clusters in low to moderate dimension, without any user-adjustable parameters. It is based on repeated tests for unimodality—using isotonic regression and a modified Hartigan dip test—applied to 1D projections of pairs of putative clusters. It handles well non-Gaussian clusters of widely varying densities and populations, and in such settings has been shown to outperform K-means variants, Gaussian mixture models, and density-based methods.

This repository contains an efficient single-threaded implementation in C++, with a MATLAB/MEX interface.

It was invented and coded by Jeremy Magland, with contributions to the algorithm and tests by Alex Barnett, at SCDA/Flatiron Institute.

Advancing Research in Basic Science and MathematicsSubscribe to Flatiron Institute announcements and other foundation updates