The DistanceMeasure instance in each DistanceMeasureCluster is used to compute a pdf() that is required by the Model interface. In order to calculate a pdf function with respect to an arbitrary vector, the distance measure used to form the cluster in the first place is needed. The ClusterClassifier uses pdf() to classify a Vector but it is not otherwise used by kmeans. The measure itself contains no state so it is not "the distance to the cluster center". It can be used; however, to calculate the distance of an arbitrary point from the cluster center as is done in pdf().
Hope this helps, Jeff -----Original Message----- From: djellel eddine Difallah [mailto:[email protected]] Sent: Friday, June 17, 2011 1:16 PM To: [email protected] Subject: Re: kmeans generates ovelapping clusters How about that measure that is now appended to the points? do you confirm that it is the distance to the cluster center? Thanks 2011/6/17 Jeff Eastman <[email protected]> > It is not unreasonable for the cluster centers to contain nonzero values > for many terms. Consider a 2-d, x-y clustering as in the DisplayKMeans > example. Every cluster center contains nonzero x and y values. The centers > themselves, taken as a whole, should be distinct and one would not expect > their 3sigma circles to overlap much. Certainly, the clustered points output > by the classification step will be assigned to a single cluster, since > kmeans is a maximum likelihood clustering algorithm. > the ouptut in my case, generated with clusterdump, put the same point in my two clusters ... > > -----Original Message----- > From: djellel eddine Difallah [mailto:[email protected]] > Sent: Friday, June 17, 2011 11:45 AM > To: [email protected] > Subject: Re: kmeans generates ovelapping clusters > > Ok. However, if the measure appended to each term has something to do with > the distance then I have found that it is different for terms that are in > both cluster. Example: > Cl1 { .... animal:0.001, ....} > Cl2 { .... animal:0.07, ....} > > What does it mean exactly ? > > 2011/6/17 Hector Yee <[email protected]> > > > One vector be a member of only one cluster but there's no requirement for > > no > > overlaps. > > You get equal radius but the cluster centers could be close enough for > them > > to overlap. > > > > On Fri, Jun 17, 2011 at 10:15 AM, djellel eddine Difallah < > > [email protected]> wrote: > > > > > Hello everyone, > > > > > > I tried kmeans on some corpus with the same script as reuters but with > -k > > 2 > > > > > > There are some terms in both generated clusters. In addition terms in a > > > cluster have a measure .. somthing like { .... animal:0.087, > > boat:3.559, > > > kitty:3.386, .....} > > > > > > Isn't kmeans supposed to generate non overlapping clusters? and what > does > > > that annotated measures mean? > > > > > > Thanks ! > > > > > > Djellel > > > > > > > > > > > -- > > Yee Yang Li Hector > > http://hectorgon.blogspot.com/ (tech + travel) > > http://hectorgon.com (book reviews) > > >
