it's almost a textbook case for unsupervised learning. Like, almost as if you could train a manifold that separated out people by genetic history (much of which was super-regionalized, and did lead to large phenotypic differences that clustered spatially)... and in that manifold "mixed-race" people would probably fall on points between two well-separated clusters.

In fact there are entire scientific papers (mainstream) doing exactly this. We've learned all sorts of interesting things, like the total genetic diversity in africa is larger than the african/eurasian diversity split.