What are some alternatives to Spherical K-Means for clustering very large datasets of high dimension?
I'm looking for something that will be fast even on large datasets, and preferably will not require you to specify the number of clusters to find.
Another property that would be nice (I'm not quite sure the name of it) is that the algorithm can calculate the clusters of a dataset A, and then later find the best cluster for a new point b which wasn't in A, without having to recalculate the clustering.