naywhayare changed the topic of #mlpack to: http://www.mlpack.org/ -- We don't respond instantly... but we will respond. Give it a few minutes. Or hours. -- Channel logs: http://www.mlpack.org/irc/
< jenkins-mlpack>
* Ryan Curtin: How did I accidentally remove two lines? I'm not actually sure.
< jenkins-mlpack>
* Ryan Curtin: First pass: make things 80 characters, minor style fixes.
< jenkins-mlpack>
* Ryan Curtin: When Conrad took the patch, he stripped out a lot of compatibility between
< jenkins-mlpack>
iterators, so... take the tests depending on that functionality out.
< jenkins-mlpack>
* Ryan Curtin: Don't add row_col_iterator support after 4.349 (currently svn trunk) since
< jenkins-mlpack>
Conrad accepted our patches.
< jenkins-mlpack>
* Ryan Curtin: Typing failure.
< jenkins-mlpack>
* Ryan Curtin: First pass: make lines 80 characters long, tabs to spaces, and bracket surgery
< jenkins-mlpack>
(or something).
< jenkins-mlpack>
* andrewmw94: X tree commit
jbc_ has joined #mlpack
< jenkins-mlpack>
Starting build #2097 for job mlpack - svn checkin test (previous build: FIXED)
< marcus_zoq>
Hello, maybe someone has an idea and can explain this; I get different signs for pca with svd and pca with eig_sym. I've compared the results with matlab and scikit only first column has another sign (using svd).
< naywhayare>
marcus_zoq: I'm not sure this is an issue; a "negative" principal component is still that same principal component
< naywhayare>
that is, it still describes the same one-dimensional subspace
< naywhayare>
or have I misunderstood?
< marcus_zoq>
No this is right, but as pointed out by ftrovato, pca and kpca with a linear kernel should have the same result, right?
govg has quit [Ping timeout: 255 seconds]
< naywhayare>
yeah, but I would say if the principal components are just negative, then it is the same result
< naywhayare>
whether or not the principal components are negative is the function of the eigendecomposition implementation, or the SVD implementation
< marcus_zoq>
matlab uses a sign convention on the coefficients "the largest element in each column will have a positive sign."
< naywhayare>
I don't think Armadillo's eigensolvers or SVD implementations make the same guarantee
< naywhayare>
although we could add code to give the same guarantee as MATLAB, I'm not sure it's necessary for correctness
< naywhayare>
where is the blue kpca-linear plot? I don't see it
< naywhayare>
maybe it's exactly the same as one of the others and gets overwritten by it?
< marcus_zoq>
yeah, KPCA (linear - naive) and matlab pca have the same values
< naywhayare>
so PCA (mlpack) appears to have a flipped axis with respect to PCA (matlab)
< naywhayare>
so I would call that, and KPCA (naive - linear) correct
< naywhayare>
but I can't say whether or not the Nystroem implementation is right. because it is sampling points, it won't necessarily give exactly the same result as PCA or naive KPCA
< marcus_zoq>
yeah, right, it is just a approximation
< naywhayare>
marcus_zoq: thanks for taking the time to respond to #361 :)