How to find relevent principal component .

Hi,

I am trying to do PCA analysis on featurevector size 30x17600, where 30 is the number of images and 17600 is the number of coffecients.

How can I find how many principal componts are requied for corect represtntaion of data.

[M N]=size(feature_vector'); m=mean(feature_vector',2); m_adj = feature_vector' - repmat(double(m),1,N); [evectors, score, evalues] = princomp(feature_vector','econ');

How to find the revelent pricomponent for multiplying with mean adujsted data

feature_vector_final=feature_vector'*evectors;

Please help me.

Thanks in advance

Answer 1

[Conrad]

Calculate the quantity:

cumsum(evalues)./sum(evalues)

The will show you the cumulative variance explained by keeping the first n components. You can also look at the following plot (called a scree plot):

xOffset = -0.2;
yOffset = 2;
nComponentsToShow = 5;

figure; hold on;
p(1) = bar(1e2*evalues(1:nComponentsToShow)/sum(evalues)); 
ylabel('Variance explained (%)');
xlabel('Factor');
set(p(1),'FaceColor','Black');
set(gca,'XTick',1:size(evalues,1));
for i = 1 : nComponentsToShow
   text(i+xOffset,1e2*evalues(i)/sum(evalues)+yOffset,sprintf...
       ('%0.2f%%',1e2*sum(evalues(1:i))/sum(evalues)));
end

Looking at the total variance explained you can decide on how many components to keep.

Conrad

[DS]

Thanks for the reply.I tried with this code.In the y-axis I am getting high amplitude and showing above the bar as 96.04% for the 1st factor for the remaining i am getting very low amplitude and showing above the bar in the range of (96 % to 100).Whether it maens only first principal componnet is required? Also I have a doubt why the xoffset = -.2 and yoffset=2; Thanks in advance