Chameleon performance evaluation
We performed all Chameleon analyses on pairwise Bray-Curtis
compositional similarities between samples (Clarke, 1993) using the
scluster function in CLUTO software version 2.1.2 (Karypis, 1999).
First, since we found little information in the literature to guide
parameter-setting, we assessed solutions of 15 clusters over a range of
neighbourhood sizes (15 – 1000 neighbours), degrees of sub-partitioning
(up to 500 sub-partitions or agglomerative phase omitted) and linkage
functions (single or complete) (Table 1). We focused carried out our
initial trials using the cluster-weighted single-link criterion function
in the agglomerative phase, as recommended for non-spherical clusters
(Karypis, 1999). For each solution we calculated average pairwise
within-cluster association (homogeneity) and the proportion of samples
located in clusters other than that of their nearest neighbour
(misclassification rate). We found using the single linkage function
caused chaining (sensu Peet & Roberts 2013) when the number of
sub-partitions specified was larger than 30. We repeated the relevant
trials using an option forcing Chameleon to prioritise large clusters
over small in the partitioning phase. On the basis of the preliminary
results, we undertook subsequent analyses using the complete linkage
function and assessed performance over a range of thematic scales (15 –
250 clusters) and degrees of sub-partitioning (30 – 500 sub-partitions)
with neighbourhood size fixed at either 30 or 1000 (Table 1).