FIGURE 7.

Results of eight unsupervised machine learning (UML) algorithms compared against Bayes Factor Delimitation (BFD*) results for eight populations of speckled dace (Rhinichthys osculus). These include four random forest (RF) methods with mixtures of classical (c) and isotonic (iso) multidimensional scaling (MDS) as well as hierarchical and partition around medoids (PAM) clustering. The t‐distributed stochastic neighbor embedding (t‐SNE) algorithm was employed with two clustering methods. Finally, a discriminant analysis of principal components (DAPC) and variational autoencoder (VAE) were applied. All UML algorithms were applied to 130 individuals genotyped at 200 SNPs. The raw number of divisions (=Clusters) for several methods included groups of individuals that shared a high proportion of missing data (*). Other cases divided individuals from two localities among clusters that did not follow any pattern (†). In one instance, Ash Meadows populations were subdivided according to springs (‡). All three scenarios were interpreted as “oversplitting” and ignored in the final interpretation (=Corrected). Lineages are as follows: PRC = Walker Sub‐basin (R. o. robustus); LVD = Long Valley; HAR = Benton Valley; ORB = Owens Valley; ASH = Ash Meadows (R. o. nevadensis); AMC = Amargosa Canyon; RFO and AMA = Oasis Valley