Self-Organizing Map example

Click the picture, please.

Figure: Map of potential Christmas presents organized by the SOM based on the data
provided by Computing Paradigms participants on 24th of November on their interests.

I edited the original eps file with Photoshop just a little
(not to be used in any serious context as such).

Command file (unix/linux, mlab)

randinit -din pjoulu3.dat -cout pjoulu.ini -topol hexa -neigh gaussian -xdim 4 -ydim 7
vsom -din pjoulu3.dat -cin pjoulu.ini -cout pjoulu.first.cod  -rlen 2000 -alpha 0.5 -radius 4
vsom -din pjoulu3.dat -cin pjoulu.first.cod -cout pjoulu.cod  -rlen 5000 -alpha 0.1 -radius 1
vcal -din pjoulu3.dat -cin pjoulu.cod -cout pjoulu.lab -numlabs 0
umat -cin pjoulu.lab -o pjoulu_umat3.eps -eps 1 -average 1 -fontsize 0.4 -W 0.5 -notitle -median
umat -cin pjoulu.lab -o pjoulu_umat3.ps -ps 1 -average 1 -fontsize 0.4 -W 0.5 -notitle -median

Command file with short explanations

Initializing the map

din: input data file
cout: initialized map
topol: topology of the map (hexa or rect)
neigh: neighborhood in the learning (gaussian or bubble) - just use either
xdim: number of map nodes in x dimension
ydim: number of map nodes in y dimension

randinit -din pjoulu3.dat -cout pjoulu.ini -topol hexa -neigh gaussian -xdim 4 -ydim 7

Teaching the map (in two phases: a) general order, b) refining)

cin: input map (already initialized)
rlen: length of the learning session
alpha: how strong changes are made (0.0 - none, 1.0 - full, i.e., rapid but possible bad quality)
radius: radius of the changes around the so called winner (quite large compared with the map size in the first round and 1 or 2 in the second) - don't neeed to bother in the first experiments if the settings are not much changed

vsom -din pjoulu3.dat -cin pjoulu.ini -cout pjoulu.first.cod  -rlen 2000 -alpha 0.5 -radius 4

vsom -din pjoulu3.dat -cin pjoulu.first.cod -cout pjoulu.cod  -rlen 5000 -alpha 0.1 -radius 1

Labeling the map

numlabs: numlabs 0 says that a node is labeled by all corresponding (best-matching) labels

vcal -din pjoulu3.dat -cin pjoulu.cod -cout pjoulu.lab -numlabs 0

Creating the map picture

umat -cin pjoulu.lab -o pjoulu_umat3.eps -eps 1 -average 1 -fontsize 0.4 -W 0.5 -notitle -median

umat -cin pjoulu.lab -o pjoulu_umat3.ps -ps 1 -average 1 -fontsize 0.4 -W 0.5 -notitle -median

More information: SOM_PAK documentation (Self-Organizing Map Program Package, Kohonen, Hynninen, Kangas, Laaksonen)

The package itself can also be downloaded (uniix or ms-dos). Note: already installed in mlab.uiah.fi.

Data file

The first row gives the number of variables. All the following rows are the actual data. Each data row consisted in this case the opinion seven persons per an item plus the name of the item in the end. The first column is thus the opinion of the first person, etc. The number 5 meant "likes a lot" (even if already has) and 1 meant "don't like at all".

7
1	2	1	2	5	2	3	suklaa
5	4	1	2	2	1	5	kylpyamme
1	1	1	5	1	1	1	nauhaton_basso
1	2	3	1	5	3	1	kermalikoori
3	2	5	5	3	5	4	adaptiivinen_patja
1	5	1	5	2	1	1	kaasuliesi
1	3	1	3	1	1	1	slov_lask.sukset
5	1	1	5	2	5	5	suom_murtsikat
1	5	5	3	4	5	1	single_malt_viski
2	3	3	1	1	1	1	norjal_vuonohevonen
2	4	5	3	3	5	3	vuosikertapunaviini
3	2	3	1	5	4	3	pelikonsoli
4	2	1	2	3	4	3	joensuul_villapaita
4	4	1	1	4	4	3	kiikarit
5	3	5	4	5	5	4	hauska_joulu
5	3	5	5	5	5	5	luminen_joulu
1	1	1	1	1	1	1	kettuturkki
1	1	3	3	3	2	1	ei_mitaan
1	4	1	2	1	1	1	tanskal_kinkku
3	4	2	1	5	5	5	etelanmatka
5	1	4	1	4	4	5	tofukalkkuna
5	2	1	5	2	4	5	riisinkeitin
3	4	1	2	1	2	1	leipakone
2	3	2	2	5	1	5	liukuri
1	4	5	5	2	3	5	espressokeitin
5	4	5	2	5	1	3	edull_vuokra-asunto
5	2	4	4	3	1	5	alykas_agentti
3	2	1	4	3	5	5	itseimuroiva_imuri
5	4	5	3	5	1	3	tytto/poikaystava
5	2	2	3	1	2	3	ilta_oopperassa
3	5	5	3	5	1	3	rakastaja/rakastajatar
5	5	5	5	5	5	5	nopea_verkkoyht_kotiin
3	5	5	1	5	3	5	videotykki
5	2	4	3	5	5	3	lumen_taju
5	1	5	1	3	5	5	lajitt_roskalaatikko
1	4	2	3	1	3	1	jouluhauki



Usage in mlab.uiah.fi

Save the commands above (or their modification) as, say, aja.run, and give:

./aja.run

in the directory in which you have the input data file.

Among the resulting files there is one eps and one ps map picture.





Timo Honkela