Haven't run it yet, but a simple change to the cost function: https://github.com/eoghanmurray/css-bipartite-compression/co... Maybe the assumption to merge more frequently than to split doesn't hold with the new cost function.