Comparison of algorithms
We compared Chameleon cluster member-sets with those derived using: i) k-means clustering; ii) flexible unweighted pair-group averaging with arithmetic mean (Belbin et al . 1992); and iii) polythetic-division (MacNaughton-Smith et al ., 1965; Belbinet al ., 1984). We transformed the adjacency matrix supplied to scluster to dissimilarity (1-simmilarity) and used each algorithm to compute solutions ranging from 15 – 250 clusters (Table 1.) We characterised each solution in terms of homogeneity and misclassification rate (as described above), the number of species occurring at higher frequencies within each cluster than in the dataset as a whole (cumulative hypergeometric probability >0.999) and the number of species with standardised phi > 0.35 (Tichy ̵́ & Chytry ̵́ 2006).