PCA should do the eigenvalue decomposition of the mean centered COVARIANCE matrix about napkinml HOT 4 CLOSED

It's possible to do PCA both by SVD and eigenvalue decomposition of the covariance matrix. Here's a reference: http://www2.imm.dtu.dk/pubdb/views/edoc_download.php/4000/pdf/imm4000

from napkinml.

What you are doing is just SVD of the mean centered data, even though there is a relationship between components of mean centered data and eigenvalues, why not calling it SVD instead?. After all , all you are doing is calling linalg.svd.

The reason i am saying this is that I assume the goal of this library is educational (if not why creating wrappers for numpy functions?), and not implementing the educational approach of PCA (check any educational resource that explains PCA on a simple way like this, or this ) , actually harms the education part of the library.

from napkinml.

Well, no that's not all that's being done... SVD is a more numerically stable way of doing PCA. If you look at other libraries most of them will use SVD instead of explicitly calculating the covariance matrix (http://scikit-learn.org/stable/modules/generated/sklearn.decomposition.PCA.html).

Here's a good explanation: https://math.stackexchange.com/questions/3869/what-is-the-intuitive-relationship-between-svd-and-pca

Also section VI in: https://arxiv.org/pdf/1404.1100.pdf

from napkinml.

@eriklindernoren There are many ways of doing the decomposition part of PCA, you could do spectral decomposition instead of SVD. SVD by itself is another common method of dimensionality reduction all by itself (quite more useful and with many other applications than just dimensionality reduction)

Anyway, it is your library, I just thought it would make more sense to provide a better implementation for people willing to check the code and learn what is going on (not just calling (the numpy implementation))

from napkinml.