Figure 1.

Algorithm for REMc. (a) REM clustering is performed on the unclustered data and then repeatedly on each new cluster until no additional subcluster is obtained. The first round of clustering yielded four clusters. Each cluster was subdivided in a second round of clustering (thick lines). Only two clusters were divided further in the third round of clustering (thin lines). No new clusters were found in the fourth round of REMc. A LL score is obtained for each round of clustering, thus when a single cluster is no longer subclassified, the LL provides a quantitative indication of the probability that the cluster represents a uniform class of data. (b) To evaluate the number of clusters obtained by REMc, EMc was performed by fixing the number of clusters between 1 and 40 (instead of determining the optimal number of clusters). The arrow indicates that the number of clusters determined by REMc, 17, was at an inflection point of the plot of LL vs cluster number.