Logs of density values can be positive because the density function can
have values > 1.
Logs of probabilities cannot be positive.
What you are looking for (i.e. cluster membership probabilities) is
given by the method
from which EM now inherits.
On Saturday, April 10, 2004, at 12:18 PM,
From: Jonathan Elsas <jelsas(a)email.unc.edu>
Date: Sat Apr 10, 2004 3:05:28 AM Pacific/Auckland
Subject: [Wekalist] How to get Probabilities from EM clusterer?
Hi -- I've been working with the EM clusterer (through Java, not
through the GUI) and am having some trouble deciphering the output.
I'd like to get the *probabilities* of cluster membership per instance
for each cluster, not the log-densities. But, I'm having a bit of
trouble coming up with those. I've been using Weka-3.4.1, and it
appears there were substantial changes in the EM class between 3.4 and
As far as I can tell, the method of the EM class that I may be
interested in is
The return values from this function seem to be all over the place --
large positive numbers, large negative numbers, or a mix. I read in
this list that the log densities should not be negative -- is that
true? if so there seems to be a bug in this method.
Is there a way to convert the output of this method from log densities
Thanks in advance,