Thanks Peter. Ure explanation helps a lot in clearing some misconceptions :) Have a good one!
Peter Perkins <> wrote in message <email@example.com>... > On 10/26/2010 2:50 AM, Alfian Abdul Halin wrote: > > I understand that to generate k-components GMMs using the > > gmdistribution.fit > > uses the k-means algorithm for cluster centroid intialization? (i was > > told this by a friend). > > Your friend is incorrect. As described in the documentation, the > default method that GMDISTRIBUTION.FIT uses to choose initial values for > the component means is a random sample from the data. You can specify > one of three methods, however. > > > > Is it possible to change that to Fuzzy C Means? > > > > I ask this because, for my data, using FCM seems to generate centroids > > that make more sense > > I'm not sure that the initial locations in this case make all that much > difference. The E-M algorithm is pretty robust with repect to initial > values, and these are _initial_ values, not the final estimates. But > you can specify starting values as one of the inputs to FIT. See the > documentation. > > Hope this helps.