[lqr]

I think the real strength of this method is that it doesn't require the data live in a vector space. Once you give up that structure, you're pretty much locked in to using points from the dataset as the cluster representatives, unless you assume some other structure.