Species complex delimitations in the genus Hedychium: A machine learning approach for cluster discovery

Statistical methods used by most morphologists to validate species boundaries (such as principal component analysis [PCA] and non-metric multidimensional scaling [nMDS]) are limiting because these methods are mostly used as visualization methods, and because the groups are identified by taxonomists (i.e., supervised), adding human bias. Here, we use a spectral clustering algorithm for the unsupervised discovery of species boundaries followed by the analysis of the cluster-defining characters.

We used spectral clustering, nMDS, and PCA on 16 morphological characters within the genus Hedychium to group 93 individuals from 10 taxa. A radial basis function kernel was used for the spectral clustering with user-specified tuning values (gamma). The goodness of the discovered clusters using each gamma value was quantified using eigengap, a normalized mutual information score, and the Rand index. Finally, mutual information-based character selection and a t-test were used to identify cluster-defining characters.

Spectral clustering revealed five, nine, and 12 clusters of taxa in the species complexes examined here. Character selection identified at least four characters that defined these clusters.

Together with our proposed character analysis methods, spectral clustering enabled the unsupervised discovery of species boundaries along with an explanation of their biological significance. Our results suggest that spectral clustering combined with a character selection analysis can enhance morphometric analyses and is superior to current clustering methods for species delimitation.