derekbeaton / gpls

a home for generalized partial least squares

Generalized Partial Least Squares

This repository is for our work on generalized partial least squares. The preprint can be found here: Beaton, D., ADNI, Saporta, G., Abdi, H. (2019). A generalization of partial least squares regression and correspondence analysis for categorical and mixed data: An application with the ADNI data. bioRxiv, 598888..

The preprint was generated through RMarkdown with the rticles package, and is fully reproducible for us at least! You can reproduce it so long as you have access to ADNI data. See the Manuscript directory for the manuscript’s content and supplements (RMarkdown, PDF).

Other materials can be found in the Code directory. A detailed RMD file provides how to reproduce the data and is found here with an HTML file.

The GPLS package lives in the Package directory. It is documented and has a dependency on the GSVD package (found here). See the Package directory for installation instructions.

gpls's Issues

Check/harmonize Y_residual

Across the plsca vs. pls approaches

PLSCA brings back the center/scale but PLS does not.
It's not a big deal and can probably stay but it would nice to either (1) verify this is consistent behavior or (2) change it so that it produces consistent behavior

Double check the discriminant example

specifically the scores & assignment. they look correct but ensure that they are also barycentric

Also give "BADA" (weighted & unweighted) a try through GPLS and check the same conditions.

rownames/colnames everywhere

they don't transfer through to all the GPLSs because of iteration.

Control for variables of no interest?

Hi Derek,

Is it possible to do something like a partial correlation (e.g., hold some variables constant) in the GPLS framework?

Thanks!

Very generic issue

Create a simple package (meant to be recycled and optmized later) that has all the core gpls algorithms and their respective techniques (e.g., PLS, CCA, RRR, PLSCA) that call into the core algos.

To determine: naming conventions vs. the use of switches, e.g., PLSC, PLSR, PLSCAN or PLS(..., algo='cor') vs. PLS(..., algo='reg'), etc...

PER USUAL: CHECK THERMOMETER

I always do this to myself.

Check the + vs. - and make sure it's correctly labeled.

Titles on existing vignettes prevent loading

I have corrected this and submitted a pull request. Do with it as you may. Thanks for putting this together Derek. -- Lynne

FI vs. FJ & FJ vs. FK

I should attempt to to harmonize the code and manuscript to reflect better names for the component scores. it's a bit of a mismatch (but easily resolved either in code or notation)