Initialization of fuzzy c-means clustering algorithm

Michael Farmer (mfarmer@utc.campus.mci.net)
Tue, 27 Feb 1996 20:06:30 +0100


I am having trouble coming up with a good way to initialize
my fuzzy c-means algorithm.

I am trying to do fuzzy c-means clustering on feature patterns
extracted from digital images. The algorithm requires either
an initial fuzzy partition (matrix of fuzzy memberships) or
some initial class centroids. My work needs to be as automated
as possible, i.e. no asking the user for single class prototypes.

Does anyone have any suggestions ? Is the common practice
to simply start off with random values that meet the constraints of
a fuzzy partition matrix ? I am not exactly thrilled with this
option myself. Also, a method that starts closer to the final
partition should speed up convergence.

Any help would be appreciated. (post or email)

Thanks,

Michael