Employing SOM after PCA

Hello,

I have applied PCA on my dataset and found out my first three variables together explained 93% of the total variance. So I decided to use only first 3 variables for SOM instead of 8 variables. I have total of 2643 observations in my dataset. There are two possible classification that I expect the SOM gives me but I don't know how to force SOM to do 2 classifications instead of 4. I use GUI for this and I would appreciate if anybody could help me with this.

Regards,

Naghmeh

Answer 1

[Greg Heath]

1. PCA ranks principal components, not original variables. So, did you deduce 3 variables from the 3 principal components?

2. PLSREGRESS ranks original variables. Use that instead.

3. SOM is an unsupervised clustering algorithm that ignores classes, IT IS NOT A CLASSIFIER!

4. PATTERNNET is a NEURAL NETWORK SUPERVISED CLASSIFIER. If you are familiar with NNs, use that.

5. Otherwise, search the STATISTICS TOOLBOX for classifiers.

6 Meanwhile, search both NEWSGROUP and ANSWERS using

 CLASSIFICATION

Hope this helps.

Thank you for formally accepting my answer

Greg

[naghmeh moradpoor]

Hi Greg,

No I did not deduce 3 variables from 3 PCs. I have managed to figure out which variables with which percentages are contributing to each PC using this code (given that I have 8 PCs and 8 variables):

******************

 for i=1:8 

    for j=8

varI = coeff(i,:)*(latent.*coeff(i,:)');

varIfromJ = coeff(i,j)*latent(j)*coeff(i,j);

percVarIfromJ = varIfromJ/varI

finalmatrix (i,8) = percVarIfromJ;

****************** My question is is it a good idea to use only those variables that contributing the most to the e.g. first three PCs in my SOM?

The output from SOM is very difficult to interpret.

Regards,

NGH

[Greg Heath]

Do not use PCA and/or SOM for supervised classification.

Greg

[Greg Heath]

However, if you have UNLABELED data and want to divide it into clusters to determine where individual classes might be, SOM and PCA + SOM are appropriate.

Hope this is clear.

Greg