My own opinion is that spectral clustering should also be able to work against a sparse connection (affinity) graph.
Anything but sparse connectivity is a complete non-starter for a scalable system. On Mon, Aug 16, 2010 at 7:00 AM, Robin Anil <[email protected]> wrote: > From a GSOC angle, it needn't be done, its upto your mentor to decide. I am > interested more in getting this completed and pushed out so that people can > really use it. If you can spare time after GSOC and still hang around the > community and help in getting this polished, it will be great. > > To create your pairwise similarity(0-1 1 means dissimilar) matrix(it can > be > the other way around as well), see the DistanceMeasure implementations. > Creating the pairwise matrix is non trivial from a scalability stand point. > > A complete spectral clustering package should take an input set of > documents, create the matrix and run clustering and output the clusters. To > get an idea of your work till now, what are the blocks missing from this > ideal package scenario? > > > Robin >
