[SciPy-Dev] Comparison of PLS algorithms

[Aman Thakral]

Whoops,  just realized, this should have gone to scikits-learn.  My mistake.

Sorry about that,
Aman

On Fri, Mar 4, 2011 at 11:34 AM, Aman Thakral <aman.thakral at gmail.com>wrote:

> Hi all,
>
> I've provided a summary of the features between the two PLS libraries
> submitted.  Please see the table below:
>
>     *Duchesnay* *Aman*  *Inner Loop* Both NIPALS and CCA NIPALS only  *Mean
> Centering* Yes Yes  *Scaling to Unit Variance* Yes Yes  *SVD alternative*
> Yes No  *Deflation of Y* Using either X or Y scores X scores Only  *Rotations
> (W* matrix)* Both X and Y X only  *Raises exceptions for improper sizing
> of input matrices* Yes No  *Provide Statistics (SPE, Hotelling's T)* No
> Yes  *Provide Statistical limits* No Yes  *Provide Variable importance to
> Projection (VIP)
> * No Yes
>
> In terms of performance, they should be same since the equations (for the
> parts that are found in both) are identical and both use numpy arrays to do
> calculations.  I think the best course of action would be to use Duchesnay's
> code for the base, then add the missing parts from my code (statistics,
> statistical limits, VIP) to help users in their analysis.  I will also get
> some examples together that uses duchesnay's code with some of the
> visualization functions to help users get a better understanding of PLS.
>
> Also, I was thinking about incorporating some inline weave for the NIPALS
> inner loop to improve the performance.  I haven't used it before, so I'll
> have figure out the inner workings of weave first.
>
> Please let me know what you think.
>
> Thanks,
> Aman
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/scipy-dev/attachments/20110304/0f4cc645/attachment.html>