I am using Weka for clustering some data and was running into a very odd problem. When I use the normal "Cluster" Tool on a data set, I am getting a result of
Cluster 1: 87 instances
Cluster 2: 88 instances
Cluster 3: 181 instances
This is what I sort of expected from the data I had, so I consider this a good result. However, I want to add this cluster as a class and save it as a new .arff file, so I am trying to use the "Add Cluster" filter that Weka provides. Now, in this filter, I select the same clusterer (EM for this experiment), and ensure that all settings are the same (even the seed number). When I apply this, I get the much worse result of
Cluster 1: 87 instances
Cluster 2: 43 instances
Cluster 3: 226 instances
This is clearly different than before, but I don't know why. I have double-checked all my settings to make sure they are identical, so I am pretty sure it is not a result of this. Also, in this question, Zannjaminderson comments on the answer that he has the same problem, so it may just not be me (I would comment to see how/if he fixed it, but alas I do not have 50 rep).
One thing that I thought could affect the result was what the clustering was trained on. In the original clustering, I used the whole training set, but for the "Add Cluster" clustering, I don't think it specifies what it trains on, so this could be the source of the error. Ultimately, my question is this: Why is this discrepancy happening and what can I do (if anything) to stop it?
