1. Find a partition of the input similarity graph (or set of points) and:
    • Split using bisection -means
    • Split using sparsest cut
  2. Recurse on each part
  3. Builds cluster-tree top-down