MCC—multiple correlation clustering

Abstract

A clustering algorithm is described which is powerful, in that at each iterative step of the method global information is used to constrain the algorithm's convergence towards a solution. It is stable in the face of missing data in the input; it is efficient in that it will extract a small signal from a lot of noise; it is impervious to multicolinearity; it may be used in two-way clustering. Each of these claims is illustrated by its application to different data sets. Despite these advantages, the algorithm is easy to implement and understand: it is sufficient to know what a correlation coefficient is in order to understand the guts of the algorithm. Because the program repeatedly correlates correlation matrices it is called here Multiple Correlation Clustering, or MCC for short.