On Mon, 6 Dec 2004, Benjamin Haibe-Kains wrote: > Hi Michael, > > I think that the differences are too important to be due to different > implementation decisions. Actually my problem is that I have a group of 1 > object and the rest in the other group when I use the 'centroid' hclust (I > use cutree to have the main two groups) and it's not the case with other > softwares. It looks like a bug in the Fortran routine but I can not access to > it. Source code is available from CRAN. In http://cran.r-project.org/src/base/R-2/R-2.0.1.tar.gz the directory R-2.0.0/src/library/stats/src contains hclust.f and hclust-utils.c. Happy debugging. ;-) > > Have you reported this "bug" before ? Can I write my 'centroid' method easily > ? > > cheers, > > benjamin > > michael watson (IAH-C) wrote: > >> Benjamin >> >> You will likely get different results from all clustering software, even >> when using the same parameters. This is because many arbitrary >> decisions have to be made during a hierarchical cluster analysis and >> different programmers will make those decisions in different ways. >> >> Mick >> >> -----Original Message----- >> From: bioconductor-bounces@stat.math.ethz.ch >> [mailto:bioconductor-bounces@stat.math.ethz.ch] On Behalf Of Benjamin >> Haibe-Kains >> Sent: 06 December 2004 11:05 >> To: Bioconductor Mailing List >> Subject: [BioC] hclust and (Eisen+ de Hoon) cluster3 program >> >> >> Hi all, >> >> I have a problem with the R function 'hclust'. I have noticed differences >> in clustering when I use the 'centroid' cluster method with 'hclust' and >> the cluster3 program (see M. Eisen and M. de Hoon). >> >> Have you noticed some differences too ? >> >> I use >> >> hclust from library 'stats' (Built: R 2.0.1; i386-pc-linux-gnu; 2004-11-15 >> 15:56:06; unix) >> cluster 3.0 using C Clustering Library version 1.25 >> >> Thanks a lot >> >> >> > > -- > Benjamin Haibe-Kains > [http://www.ulb.ac.be/di/map/bhaibeka/] > > > PhD student in the Machine Learning Group (MLG) > [http://www.ulb.ac.be/di/mlg/] > Universite Libre de Bruxelles (ULB) > E-mail: bhaibeka@ulb.ac.be > MicroArray Unity [http://www.bordet.be/servmed/array/index.htm] > Institut Jules Bordet (IJB) > E-mail: benjamin.haibekains@bordet.be > > _______________________________________________ > Bioconductor mailing list > Bioconductor@stat.math.ethz.ch > https://stat.ethz.ch/mailman/listinfo/bioconductor > > Charles C. Berry (858) 534-2098 Dept of Family/Preventive Medicine E mailto:cberry@tajo.ucsd.edu UC San Diego http://hacuna.ucsd.edu/members/ccb.html La Jolla, San Diego 92093-0717

Dear all, I saw (be maybe on a older version of Eisen software) a problem of precision, I sent him this remark (Apr 2002): --- Old Message --- I used simple data (see below) to understand the hierarchical clustering, and I did find the same results with Maple (not very convenient !) but with a very different precision. Example (Distance: correlation centered, average link): NODE1X GENE15X GENE10X 0.9996337890625 NODE2X GENE20X GENE16X 0.99957275390625 NODE3X GENE14X GENE11X 0.99835205078125 Maple: v Node1: .99959179339780276201 Node2: .99956936766825333998 Node3: .99833748845958267738 I thought that Cluster use Double precision, but it should have something like 15 good digits. Fortunately, data were very short, and with the same order of magnitude, but a computer scientist told me that floating point precision is far more less if operands (in addition, substraction...) differ greatly in size. -------------- Data: UNIQID NAME GWEIGHT GORDER "V1" "V2" "V3" EWEIGHT 1 1 1 "A1" 1 1 2 16 18 "A2" 1 2 12 9 7 "A3" 1 3 9 10 4 "A4" 1 4 5 2 12 "A5" 1 5 12 14 7 "A6" 1 6 9 16 10 "A7" 1 7 8 10 10 "A8" 1 8 10 6 6 "A9" 1 9 14 1 28 "A10" 1 10 9 10 23 "A11" 1 11 9 16 27 "A12" 1 12 17 12 37 "A13" 1 13 15 5 23 "A14" 1 14 7 14 29 "A15" 1 15 11 8 29 "A16" 1 16 4 16 37 "A17" 1 17 32 25 34 "A18" 1 18 28 35 30 "A19" 1 19 30 28 23 "A20" 1 20 32 22 28 "A21" 1 21 25 22 26 "A22" 1 22 27 33 26 "A23" 1 23 28 33 31 "A24" 1 24 36 28 31 --- On Mon, 06 Dec 2004 13:09:46 +0100 Benjamin Haibe-Kains <bhaibeka@ulb.ac.be> wrote: > Hi Michael, > > I think that the differences are too important to be due to different > implementation decisions. Actually my problem is that I have a group of > 1 object and the rest in the other group when I use the 'centroid' > hclust (I use cutree to have the main two groups) and it's not the case > with other softwares. It looks like a bug in the Fortran routine but I > can not access to it. > > Have you reported this "bug" before ? Can I write my 'centroid' method > easily ? > > cheers, > > benjamin > > michael watson (IAH-C) wrote: > > >Benjamin > > > >You will likely get different results from all clustering software, even > >when using the same parameters. This is because many arbitrary > >decisions have to be made during a hierarchical cluster analysis and > >different programmers will make those decisions in different ways. > > > >Mick > > > >-----Original Message----- > >From: bioconductor-bounces@stat.math.ethz.ch > >[mailto:bioconductor-bounces@stat.math.ethz.ch] On Behalf Of Benjamin > >Haibe-Kains > >Sent: 06 December 2004 11:05 > >To: Bioconductor Mailing List > >Subject: [BioC] hclust and (Eisen+ de Hoon) cluster3 program > > > > > >Hi all, > > > >I have a problem with the R function 'hclust'. I have noticed > >differences in clustering when I use the 'centroid' cluster method with > >'hclust' and the cluster3 program (see M. Eisen and M. de Hoon). > > > >Have you noticed some differences too ? > > > >I use > > > >hclust from library 'stats' (Built: R 2.0.1; i386-pc-linux-gnu; > >2004-11-15 15:56:06; unix) > >cluster 3.0 using C Clustering Library version 1.25 > > > >Thanks a lot > > > > > > > -- Antoine Lucas Centre de g?n?tique Mol?culaire, CNRS 91198 Gif sur Yvette Cedex Tel: (33)1 69 82 38 89 Fax: (33)1 69 82 38 77