This is a faster version of an algorithm I introduced a while back, but it turns out it doesn’t help solve anything new, and it’s not as fast as my sort-based algorithms, so I’m not including it in Black Tree Massive. The idea is that human generated classifier labels might be to some extent arbitrary, even though they define real world classes. Specifically, that individual classes actually have sub-classes that share a top-level human generated classifier label. This algorithm finds the “natural” classes on an unsupervised basis, and uses those for prediction. The prediction step is in the previous article.