Unsupervised Clustering based entirely on entity similarity

Question

I'm trying to cluster a number of entities. The only metric I have for clustering is the similarity between one entity and another. (a 0-1 float)

Can anyone think of any clustering algorithms that would achieve this? I appreciate it would be returning fuzzy groups. That's still a better solution than estimating groups programatically as I am at the moment.

Answer 1

All clustering algorithms that act on networks (also known as graphs - nodes connected by edges) that take into account neighbourhood topology. Many of these exist. I recommend RNSC and MCL (disclaimer: I wrote the latter). To elaborate, cluster structure in simple (unweighted) networks can exist in the form of node sets where many of the possible connections are realised.

Answer 2

You can use DBSCAN (Wikipedia) or OPTICS (Wikipedia). While they talk about distances, they do not require metric properties or anything - they can be used with similarity functions just as well!