No easy answer. It is true that the inclusion of many noisy variables can make it difficult to see clusters or other interesting trends in the data. On the other hand, methods that eliminate variables univariately are not able to detect that a variable may be very useful when studied in combination with other variables. I think it is difficult to eliminate irrelevant variables unless you know what the answer is supposed to be.
----------------------- F. James Rohlf, Distinguished Professor & Graduate Program Director State University of New York, Stony Brook, NY 11794-5245 www: http://life.bio.sunysb.edu/ee/rohlf > -----Original Message----- > From: morphmet [mailto:[EMAIL PROTECTED] > Sent: Friday, January 13, 2006 4:05 PM > To: morphmet > Subject: Variable selection in clustering > > Hello all, > > I wonder if someone might help with advice on approaches to > selecting variables when clustering cases (I'm using several > methods - Ward, bagged k-means, etc) - and am working with a > large number of apparently relevant variables. I fear that > "noisy" or irrelevant variables may be weakening my analysis > and I would like to refine the input space by identifying and > then eliminating any nuisance variables. > > My concern is to locate procedures (hopefully software) to > select a best sub-set of variables as input; and then refine > the input space following my initial exploratory clustering. > I can use discriminant analysis or simply examine univariate > F ratios - but would would seem to simply bias any subsequent > runs towards the classification structure produced by the > first analysis? > > Are there any other procedures for estimating the relative > power of the input variables? and then refining the input space? > > Regards > > Tim Brennan > -- > Replies will be sent to the list. > For more information visit http://www.morphometrics.org > -- Replies will be sent to the list. For more information visit http://www.morphometrics.org
