Closest match I can get is to run:
data=rand(100,10); % data set
[W,pc] = pca(cov(data));
then don't demean
data2=data
[W2, EvalueMatrix2] = eig(cov(data2));
[W3, EvalueMatrix3] = svd(cov(data2));
In this case W2 and W3 agree and W is the transpose of them?
Still not clear why W should be the transpose of the other two?
As an extra check I use pcacov:
[W4, EvalueMatrix4] = pcacov(cov(data2));
Again it agrees with WE and W3 but is the transpose of W?
The results are different because you're subtracting the mean of each row of the data matrix. Based on the way you're computing things, rows of the data matrix correspond to data points and columns correspond to dimensions (this is how the pca() function works too). With this setup, you should subtract the mean from each column, not row. This corresponds to 'centering' the data; the mean along each dimension is set to zero. Once you do this, results should be equivalent to pca(), up to sign flips.
Edit to address edited question:
Centering issue looks ok now. When you run the eigenvalue decomposition on the covariance matrix, remember to sort the eigenvectors in order of descending eigenvalues. This should match the output of pcacov(). When calling pca(), you have to pass it the data matrix, not the covariance matrix.
Related
Suppose there is a matrix B, where its size is a 500*1000 double(Here, 500 represents the number of observations and 1000 represents the number of features).
sigma is the covariance matrix of B, and D is a diagonal matrix whose diagonal elements are the eigenvalues of sigma. Assume A is the eigenvectors of the covariance matrix sigma.
I have the following questions:
I need to select the first k = 800 eigenvectors corresponding to the eigenvalues with the largest magnitude to rank the selected features. The final matrix named Aq. How can I do this in MATLAB?
What is the meaning of these selected eigenvectors?
It seems the size of the final matrix Aq is 1000*800 double once I calculate Aq. The time points/observation information of 500 has disappeared. For the final matrix Aq, what does the value 1000 in matrix Aq represent now? Also, what does the value 800 in matrix Aq represent now?
I'm assuming you determined the eigenvectors from the eig function. What I would recommend to you in the future is to use the eigs function. This not only computes the eigenvalues and eigenvectors for you, but it will compute the k largest eigenvalues with their associated eigenvectors for you. This may save computational overhead where you don't have to compute all of the eigenvalues and associated eigenvectors of your matrix as you only want a subset. You simply supply the covariance matrix of your data to eigs and it returns the k largest eigenvalues and eigenvectors for you.
Now, back to your problem, what you are describing is ultimately Principal Component Analysis. The mechanics behind this would be to compute the covariance matrix of your data and find the eigenvalues and eigenvectors of the computed result. It has been known that doing it this way is not recommended due to numerical instability with computing the eigenvalues and eigenvectors for large matrices. The most canonical way to do this now is via Singular Value Decomposition. Concretely, the columns of the V matrix give you the eigenvectors of the covariance matrix, or the principal components, and the associated eigenvalues are the square root of the singular values produced in the diagonals of the matrix S.
See this informative post on Cross Validated as to why this is preferred:
https://stats.stackexchange.com/questions/79043/why-pca-of-data-by-means-of-svd-of-the-data
I'll throw in another link as well that talks about the theory behind why the Singular Value Decomposition is used in Principal Component Analysis:
https://stats.stackexchange.com/questions/134282/relationship-between-svd-and-pca-how-to-use-svd-to-perform-pca
Now let's answer your question one at a time.
Question #1
MATLAB generates the eigenvalues and the corresponding ordering of the eigenvectors in such a way where they are unsorted. If you wish to select out the largest k eigenvalues and associated eigenvectors given the output of eig (800 in your example), you'll need to sort the eigenvalues in descending order, then rearrange the columns of the eigenvector matrix produced from eig then select out the first k values.
I should also note that using eigs will not guarantee sorted order, so you will have to explicitly sort these too when it comes down to it.
In MATLAB, doing what we described above would look something like this:
sigma = cov(B);
[A,D] = eig(sigma);
vals = diag(D);
[~,ind] = sort(abs(vals), 'descend');
Asort = A(:,ind);
It's a good thing to note that you do the sorting on the absolute value of the eigenvalues because scaled eigenvalues are also eigenvalues themselves. These scales also include negatives. This means that if we had a component whose eigenvalue was, say -10000, this is a very good indication that this component has some significant meaning to your data, and if we sorted purely on the numbers themselves, this gets placed near the lower ranks.
The first line of code finds the covariance matrix of B, even though you said it's already stored in sigma, but let's make this reproducible. Next, we find the eigenvalues of your covariance matrix and the associated eigenvectors. Take note that each column of the eigenvector matrix A represents one eigenvector. Specifically, the ith column / eigenvector of A corresponds to the ith eigenvalue seen in D.
However, the eigenvalues are in a diagonal matrix, so we extract out the diagonals with the diag command, sort them and figure out their ordering, then rearrange A to respect this ordering. I use the second output of sort because it tells you the position of where each value in the unsorted result would appear in the sorted result. This is the ordering we need to rearrange the columns of the eigenvector matrix A. It's imperative that you choose 'descend' as the flag so that the largest eigenvalue and associated eigenvector appear first, just like we talked about before.
You can then pluck out the first k largest vectors and values via:
k = 800;
Aq = Asort(:,1:k);
Question #2
It's a well known fact that the eigenvectors of the covariance matrix are equal to the principal components. Concretely, the first principal component (i.e. the largest eigenvector and associated largest eigenvalue) gives you the direction of the maximum variability in your data. Each principal component after that gives you variability of a decreasing nature. It's also good to note that each principal component is orthogonal to each other.
Here's a good example from Wikipedia for two dimensional data:
I pulled the above image from the Wikipedia article on Principal Component Analysis, which I linked you to above. This is a scatter plot of samples that are distributed according to a bivariate Gaussian distribution centred at (1,3) with a standard deviation of 3 in roughly the (0.878, 0.478) direction and of 1 in the orthogonal direction. The component with a standard deviation of 3 is the first principal component while the one that is orthogonal is the second component. The vectors shown are the eigenvectors of the covariance matrix scaled by the square root of the corresponding eigenvalue, and shifted so their tails are at the mean.
Now let's get back to your question. The reason why we take a look at the k largest eigenvalues is a way of performing dimensionality reduction. Essentially, you would be performing a data compression where you would take your higher dimensional data and project them onto a lower dimensional space. The more principal components you include in your projection, the more it will resemble the original data. It actually begins to taper off at a certain point, but the first few principal components allow you to faithfully reconstruct your data for the most part.
A great visual example of performing PCA (or SVD rather) and data reconstruction is found by this great Quora post I stumbled upon in the past.
http://qr.ae/RAEU8a
Question #3
You would use this matrix to reproject your higher dimensional data onto a lower dimensional space. The number of rows being 1000 is still there, which means that there were originally 1000 features in your dataset. The 800 is what the reduced dimensionality of your data would be. Consider this matrix as a transformation from the original dimensionality of a feature (1000) down to its reduced dimensionality (800).
You would then use this matrix in conjunction with reconstructing what the original data was. Concretely, this would give you an approximation of what the original data looked like with the least amount of error. In this case, you don't need to use all of the principal components (i.e. just the k largest vectors) and you can create an approximation of your data with less information than what you had before.
How you reconstruct your data is very simple. Let's talk about the forward and reverse operations first with the full data. The forward operation is to take your original data and reproject it but instead of the lower dimensionality, we will use all of the components. You first need to have your original data but mean subtracted:
Bm = bsxfun(#minus, B, mean(B,1));
Bm will produce a matrix where each feature of every sample is mean subtracted. bsxfun allows the subtraction of two matrices in unequal dimension provided that you can broadcast the dimensions so that they can both match up. Specifically, what will happen in this case is that the mean of each column / feature of B will be computed and a temporary replicated matrix will be produced that is as large as B. When you subtract your original data with this replicated matrix, the effect will subtract every data point with their respective feature means, thus decentralizing your data so that the mean of each feature is 0.
Once you do this, the operation to project is simply:
Bproject = Bm*Asort;
The above operation is quite simple. What you are doing is expressing each sample's feature as a linear combination of principal components. For example, given the first sample or first row of the decentralized data, the first sample's feature in the projected domain is a dot product of the row vector that pertains to the entire sample and the first principal component which is a column vector.. The first sample's second feature in the projected domain is a weighted sum of the entire sample and the second component. You would repeat this for all samples and all principal components. In effect, you are reprojecting the data so that it is with respect to the principal components - which are orthogonal basis vectors that transform your data from one representation to another.
A better description of what I just talked about can be found here. Look at Amro's answer:
Matlab Principal Component Analysis (eigenvalues order)
Now to go backwards, you simply do the inverse operation, but a special property with the eigenvector matrix is that if you transpose this, you get the inverse. To get the original data back, you undo the operation above and add the means back to the problem:
out = bsxfun(#plus, Bproject*Asort.', mean(B, 1));
You want to get the original data back, so you're solving for Bm with respect to the previous operation that I did. However, the inverse of Asort is just the transpose here. What's happening after you perform this operation is that you are getting the original data back, but the data is still decentralized. To get the original data back, you must add the means of each feature back into the data matrix to get the final result. That's why we're using another bsxfun call here so that you can do this for each sample's feature values.
You should be able to go back and forth from the original domain and projected domain with the above two lines of code. Now where the dimensionality reduction (or the approximation of the original data) comes into play is the reverse operation. What you need to do first is project the data onto the bases of the principal components (i.e. the forward operation), but now to go back to the original domain where we are trying to reconstruct the data with a reduced number of principal components, you simply replace Asort in the above code with Aq and also reduce the amount of features you're using in Bproject. Concretely:
out = bsxfun(#plus, Bproject(:,1:k)*Aq.', mean(B, 1));
Doing Bproject(:,1:k) selects out the k features in the projected domain of your data, corresponding to the k largest eigenvectors. Interestingly, if you just want the representation of the data with regards to a reduced dimensionality, you can just use Bproject(:,1:k) and that'll be enough. However, if you want to go forward and compute an approximation of the original data, we need to compute the reverse step. The above code is simply what we had before with the full dimensionality of your data, but we use Aq as well as selecting out the k features in Bproject. This will give you the original data that is represented by the k largest eigenvectors / eigenvalues in your matrix.
If you'd like to see an awesome example, I'll mimic the Quora post that I linked to you but using another image. Consider doing this with a grayscale image where each row is a "sample" and each column is a feature. Let's take the cameraman image that's part of the image processing toolbox:
im = imread('camerman.tif');
imshow(im); %// Using the image processing toolbox
We get this image:
This is a 256 x 256 image, which means that we have 256 data points and each point has 256 features. What I'm going to do is convert the image to double for precision in computing the covariance matrix. Now what I'm going to do is repeat the above code, but incrementally increasing k at each go from 3, 11, 15, 25, 45, 65 and 125. Therefore, for each k, we are introducing more principal components and we should slowly start to get a reconstruction of our data.
Here's some runnable code that illustrates my point:
%%%%%%%// Pre-processing stage
clear all;
close all;
%// Read in image - make sure we cast to double
B = double(imread('cameraman.tif'));
%// Calculate covariance matrix
sigma = cov(B);
%// Find eigenvalues and eigenvectors of the covariance matrix
[A,D] = eig(sigma);
vals = diag(D);
%// Sort their eigenvalues
[~,ind] = sort(abs(vals), 'descend');
%// Rearrange eigenvectors
Asort = A(:,ind);
%// Find mean subtracted data
Bm = bsxfun(#minus, B, mean(B,1));
%// Reproject data onto principal components
Bproject = Bm*Asort;
%%%%%%%// Begin reconstruction logic
figure;
counter = 1;
for k = [3 11 15 25 45 65 125 155]
%// Extract out highest k eigenvectors
Aq = Asort(:,1:k);
%// Project back onto original domain
out = bsxfun(#plus, Bproject(:,1:k)*Aq.', mean(B, 1));
%// Place projection onto right slot and show the image
subplot(4, 2, counter);
counter = counter + 1;
imshow(out,[]);
title(['k = ' num2str(k)]);
end
As you can see, the majority of the code is the same from what we have seen. What's different is that I loop over all values of k, project back onto the original space (i.e. computing the approximation) with the k highest eigenvectors, then show the image.
We get this nice figure:
As you can see, starting with k=3 doesn't really do us any favours... we can see some general structure, but it wouldn't hurt to add more in. As we start increasing the number of components, we start to get a clearer picture of what the original data looks like. At k=25, we actually can see what the cameraman looks like perfectly, and we don't need components 26 and beyond to see what's happening. This is what I was talking about with regards to data compression where you don't need to work on all of the principal components to get a clear picture of what's going on.
I'd like to end this note by referring you to Chris Taylor's wonderful exposition on the topic of Principal Components Analysis, with code, graphs and a great explanation to boot! This is where I got started on PCA, but the Quora post is what solidified my knowledge.
Matlab - PCA analysis and reconstruction of multi dimensional data
Suppose I have matrix A of 100x200x300. Third dimension is called "page" in Matlab and this matrix has 300 pages then.
Now I want to calculate standard deviation within each page and get a result matrix of 1x1x300.
I can't just do
std(std(A,0,1),0,2)
because normalization will be wrong as I think.
You need to collapse the first two dimensions into one (i.e. into columns) using reshape; and then compute std along each column:
Ar = reshape(A, size(A,1)*size(A,2), size(A,3));
result = std(Ar);
This will give you a 1x300 vector as result. If you really need it to be 1x1x300, use
result = shiftdim(result, -1);
Please see the following issue:
P=rand(4,4);
for i=1:size(P,2)
for j=1:size(P,2)
[r,p]=corr(P(:,i),P(:,j))
end
end
Clearly, the loop will cause the number of correlations to be doubled (i.e., corr(P(:,1),P(:,4)) and corr(P(:,4),P(:,1)). Does anyone have a suggestion on how to avoid this? Perhaps not using a loop?
Thanks!
I have four suggestions for you, depending on what exactly you are doing to compute your matrices. I'm assuming the example you gave is a simplified version of what needs to be done.
First Method - Adjusting the inner loop index
One thing you can do is change your j loop index so that it only goes from 1 up to i. This way, you get a lower triangular matrix and just concentrate on the values within the lower triangular half of your matrix. The upper half would essentially be all set to zero. In other words:
for i = 1 : size(P,2)
for j = 1 : i
%// Your code here
end
end
Second Method - Leave it unchanged, but then use unique
You can go ahead and use the same matrix like you did before with the full two for loops, but you can then filter the duplicates by using unique. In other words, you can do this:
[Y,indices] = unique(P);
Y will give you a list of unique values within the matrix P and indices will give you the locations of where these occurred within P. Note that these are column major indices, and so if you wanted to find the row and column locations of where these locations occur, you can do:
[rows,cols] = ind2sub(size(P), indices);
Third Method - Use pdist and squareform
Since you're looking for a solution that requires no loops, take a look at the pdist function. Given a M x N matrix, pdist will find distances between each pair of rows in a matrix. squareform will then transform these distances into a matrix like what you have seen above. In other words, do this:
dists = pdist(P.', 'correlation');
distMatrix = squareform(dists);
Fourth Method - Use the corr method straight out of the box
You can just use corr in the following way:
[rho, pvals] = corr(P);
corr in this case will produce a m x m matrix that contains the correlation coefficient between each pair of columns an n x m matrix stored in P.
Hopefully one of these will work!
this works ?
for i=1:size(P,2)
for j=1:i
Since you are just correlating each column with the other, then why not just use (straight from the documentation)
[Rho,Pval] = corr(P);
I don't have the Statistics Toolbox, but according to http://www.mathworks.com/help/stats/corr.html,
corr(X) returns a p-by-p matrix containing the pairwise linear correlation coefficient between each pair of columns in the n-by-p matrix X.
I try to calculate the correlation matrix of a set of histogram vectors. But the result is a truncated version of what I (think) I want. I have 200 histograms by 32 bins each. The result from
correlation_matrix = corrcoef(set_of_histograms)
is a 32 by 32 matrix.
I want to use this to calculate how my original histograms match up. (this by later using eigs and other stuff).
But which correlation method is right for this? I have tried "corrcoef" but there are "corr" and "cov" as well. Can't understand their differences by reading matlab help...
correlation_matrix = corrcoef(set_of_histograms')
(Note the ')
1) corrcoef treats every column as an observation, and calculates the correlations between each pair. I'm assuming your histograms matrix is 200x32; hence, in your case, every row is an observation. If you transpose your histograms matrix before running corrcoef, you should get the 200x200 result you're looking for:
[rho, p] = corrcoef( set_of_histograms' );
(' transposes the matrix)
2) cov returns the covariance matrix, not the correlation; while the covariance matrix is used in calculating the correlation, it is not the measure you're looking for.
3) As for corr and corrcoef, they have a few implementation differences between them. As long as you are only interested in Pearson's correlation, they are identical for your purposes. corr also has an option to calculate Spearman's or Kendall's correlations, which corrcoef does not have.
I want to use the "princomp" function of Matlab but this function gives the eigenvalues in a sorted array. This way I can't find out to which column corresponds which eigenvalue.
For Matlab,
m = [1,2,3;4,5,6;7,8,9];
[pc,score,latent] = princomp(m);
is the same as
m = [2,1,3;5,4,6;8,7,9];
[pc,score,latent] = princomp(m);
That is, swapping the first two columns does not change anything. The result (eigenvalues) in latent will be: (27,0,0)
The information (which eigenvalue corresponds to which original (input) column) is lost.
Is there a way to tell matlab to not to sort the eigenvalues?
With PCA, each principle component returned will be a linear combination of the original columns/dimensions. Perhaps an example might clear up any misunderstanding you have.
Lets consider the Fisher-Iris dataset comprising of 150 instances and 4 dimensions, and apply PCA on the data. To make things easier to understand, I am first zero-centering the data before calling PCA function:
load fisheriris
X = bsxfun(#minus, meas, mean(meas)); %# so that mean(X) is the zero vector
[PC score latent] = princomp(X);
Lets look at the first returned principal component (1st column of PC matrix):
>> PC(:,1)
0.36139
-0.084523
0.85667
0.35829
This is expressed as a linear combination of the original dimensions, i.e.:
PC1 = 0.36139*dim1 + -0.084523*dim2 + 0.85667*dim3 + 0.35829*dim4
Therefore to express the same data in the new coordinates system formed by the principal components, the new first dimension should be a linear combination of the original ones according to the above formula.
We can compute this simply as X*PC which is the exactly what is returned in the second output of PRINCOMP (score), to confirm this try:
>> all(all( abs(X*PC - score) < 1e-10 ))
1
Finally the importance of each principal component can be determined by how much variance of the data it explains. This is returned by the third output of PRINCOMP (latent).
We can compute the PCA of the data ourselves without using PRINCOMP:
[V E] = eig( cov(X) );
[E order] = sort(diag(E), 'descend');
V = V(:,order);
the eigenvectors of the covariance matrix V are the principal components (same as PC above, although the sign can be inverted), and the corresponding eigenvalues E represent the amount of variance explained (same as latent). Note that it is customary to sort the principal component by their eigenvalues. And as before, to express the data in the new coordinates, we simply compute X*V (should be the same as score above, if you make sure to match the signs)
"The information (which eigenvalue corresponds to which original (input) column) is lost."
Since each principal component is a linear function of all input variables, each principal component (eigenvector, eigenvalue), corresponds to all of the original input columns. Ignoring possible changes in sign, which are arbitrary in PCA, re-ordering the input variables about will not change the PCA results.
"Is there a way to tell matlab to not to sort the eigenvalues?"
I doubt it: PCA (and eigen analysis in general) conventionally sorts the results by variance, though I'd note that princomp() sorts from greatest to least variance, while eig() sorts in the opposite direction.
For more explanation of PCA using MATLAB illustrations, with or without princomp(), see:
Principal Components Analysis