[Eeglablist] Silhouette value for IC clustering

[時本真吾]

Dear Makoto,

Thank you for the quick and kind reply. This is really practical and easy to understand. Is this method published anywhere? I am referring to Takahashi & Kitazawa (2017, JNS) for Silhouette value as a criterion of IC clustering.

******************************************
Shingo Tokimoto, Ph.D.
in Linguistics and Psychology
Department of Foreign Languages
Mejiro University
4-31-1, Naka-Ochiai, Shinjuku, Tokyo,
161-8539, Japan
tokimoto at mejiro.ac.jp
******************************************

> 2017/11/29 3:04、Makoto Miyakoshi <mmiyakoshi at ucsd.edu>のメール:
> 
> Dear Shingo,
> 
> Another practical criterion to determine the number of clusters is to specify the minimum number of unique subjects per cluster. For example, if you determine to include at least 70% of unique subjects, after clustering you count how many unique subjects are included, and if it did not reach 70%, you lower the number of clusters and try it again until it meets the 70%-criterion. In most of my plugins for STUDY anlaysis, it is reported how many unique subjects are included by each cluster.
> 
> Makoto
> 
> On Sat, Nov 25, 2017 at 2:12 AM, 時本真吾 <tokimoto at mejiro.ac.jp> wrote:
> Dear Dr. Jens Bernhardsson,
> 
> I am deeply grateful for the excellent code. I am very sorry for the late response. This is because I took some time to understand the meaning of plotted results. I am trying to cluster 323 ICs from 31 participants by dipoles. I repeated the code 10  times, and the number of clusters for which the average Silhouette value was greatest was 15 or 16. Can I understand that the best number of clusters here is 15 or 16? Thank you very much again.
> 
> ******************************************
> Shingo Tokimoto, Ph.D.
> in Linguistics and Psychology
> Department of Foreign Languages
> Mejiro University
> 4-31-1, Naka-Ochiai, Shinjuku, Tokyo,
> 161-8539, Japan
> tokimoto at mejiro.ac.jp
> ******************************************
> 
> > 2017/11/17 19:19、Bernhardsson Jens <Jens.Bernhardsson at miun.se>のメール:
> >
> > Hi,
> >
> > I believe that the Silhouette value found in A10: MI-clust is based on kmeans cluster of multidimensional scaled data performed on the mutual information matrix and the kmeans found in STUDY is based on a pca derived (pre clustering) array. You can run the code below after you pre cluster in STUDY to get the mean and individual Silhouette coefficients for different cluster solutions.
> >
> > avgSil = [];
> > kmod = [];
> > siSil = [];
> >
> > for num_of_cluster = 2:25
> >    kmod = kmeans(STUDY.etc.preclust.preclustdata,num_of_cluster, 'replicates',10,'emptyaction','drop','distance','sqEuclidean');
> >    siSil(:,num_of_cluster-1) = silhouette(STUDY.etc.preclust.preclustdata,kmod,'sqEuclidean');
> >    avgSil = [avgSil; num_of_cluster mean(siSil(:,num_of_cluster-1))];
> > end
> >
> > figure;
> > plot( avgSil(:,1),avgSil(:,2),'r*-.');
> > % ylim([0 1])
> > set(gca,'XTick',1:num_of_cluster, 'XGrid', 'on');
> >
> >
> > Regards,
> > Jens
> >
> > -----Ursprungligt meddelande-----
> > Från: eeglablist [mailto:eeglablist-bounces at sccn.ucsd.edu] För ????
> > Skickat: den 17 september 2017 15:43
> > Till: eeglablist at sccn.ucsd.edu
> > Ämne: [Eeglablist] Silhouette value for IC clustering
> >
> > Dear EEGLAB users,
> >
> > I understand that the determination of the number of clusters in IC clustering is one of the difficult problems in source localizations by ICs. Silhouette value is one of the criteria for the best number of clusters in IC clustering, and I have found the description of silhouette value in EEGLAB Wiki (A10: MI-clust) for a single participant. Can we get silhouette values for group data in STUDY formats? Thank you in advance.
> >
> > ******************************************
> > Shingo Tokimoto, Ph.D.
> > in Linguistics and Psychology
> > Department of Foreign Languages
> > Mejiro University
> > 4-31-1, Naka-Ochiai, Shinjuku, Tokyo,
> > 161-8539, Japan
> > tokimoto at mejiro.ac.jp
> > ******************************************
> >
> > _______________________________________________
> > Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
> > To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu
> > For digest mode, send an email with the subject "set digest mime" to eeglablist-request at sccn.ucsd.edu
> 
> _______________________________________________
> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu
> For digest mode, send an email with the subject "set digest mime" to eeglablist-request at sccn.ucsd.edu
> 
> 
> 
> -- 
> Makoto Miyakoshi
> Swartz Center for Computational Neuroscience
> Institute for Neural Computation, University of California San Diego