Clustering noisy data with deterministic annealing
A cluster algorithm for noisy data distributions is presented. It minimizes an error function using a deterministic annealing procedure. Phase transitions occur during the annealing as large clusters split into smaller ones. Critical ``temperatures'' corresponding to these transitions are estimated in order to make the annealing as efficient as possible. The approach is successfully tested on data sets containing up to 10 clusters contaminated with 100% noise.
LU TP 93-28