tag:blogger.com,1999:blog-6555947.post3721237711636351029..comments2023-03-24T07:33:48.282-06:00Comments on The Geomblog: Choosing the number of clusters II: Diminishing Returns and the ROC curveSuresh Venkatasubramanianhttp://www.blogger.com/profile/15898357513326041822noreply@blogger.comBlogger2125tag:blogger.com,1999:blog-6555947.post-43705359443875151802010-03-23T22:13:16.438-06:002010-03-23T22:13:16.438-06:00Diminishing returns idea sounds interesting. I hav...Diminishing returns idea sounds interesting. I have to look into it more closely. <br /><br />I am under impression that with information theory we can only get well grounded solution to the number of clusters problem. With the interpretation that clustering models the data and so total cost is the data encoded with the model and encoded model. So minimizing that total cost should give us best fit clustering model given original model assumptions (like spherical clusters in the k-means criterion). <br /><br />Is there other possible interpretation? Elbow idea is ok also (and it is known to work ok). F-ratio statistic is a similar idea. But I don't see how it would be theoretically justified. <br /><br />Ah, then there is idea of clustering series of random sampled subsets of the original data. Clusterings that are similar with some fixed k are assumed to be correct. Maybe this could be called cross-validation method. But it is a little bit too statistical to my liking. :-)<br /><br />I am actually acutely interested in this now, as one of my current projects is to give solution of speaker clustering problem in speaker diarization problem. In speaker diarization we are given a meeting recording with unknown speakers and task is to segment the recording according to speakers (giving answer who speaks and when). Current solutions are fairly ad-hoc.Unknownhttps://www.blogger.com/profile/00227991895773060923noreply@blogger.comtag:blogger.com,1999:blog-6555947.post-9325585316762065352010-03-13T13:06:10.966-07:002010-03-13T13:06:10.966-07:00I think we should come up with some other food pro...I think we should come up with some other food processes as well -- the New York Deli? The Sushi Boat?Anonymousnoreply@blogger.com