Clustering

[trevorsummerssmith]

Eric fill out some random thoughts.

[epurdy]

The main difficulty with clustering is figuring out an intelligible representation of a cluster. We want to be able to look at a cluster that contains maybe 25% of all the vertices, and have some idea what its "deal" is.

This basically means having some sort of domain-specific "summarization" operators.

OR: this is a weirder idea, but you could try "summarizing by sampling": you show a bunch of random examples from a given cluster. Then you can be pretty sure that the intelligible clusters will "look right" most of the time. Unintelligible clusters will probably at least look unintelligible, because the user won't be able to detect any sort of pattern from the random examples shown.