[snovv_crash]

It's boring, in a sense that it always gives reasonable results and is easy to implement. It also scales well in N, D and K, and from my experience converges in just a few iterations from anything better than a pure random initialisation strategy.

IMO it is very good as a final clustering algorithm once you've already applied some more complex transformations on your data to linearise it and account for deeper knowledge of the problem. This might be a spectral space transformation (you care about connectedness), or an embedding (you care about whatever the network was trained on) or descriptor (you care about the algorithm's similarity).

But once you've applied the transform, you then have a minimalist fast scalable clustering that just does clustering and doesn't need to know anything more about the problem being solved. Very unix-y feeling.