[WIP] [R] Replace vignettes and examples

[david-cortes]

ref #9810
closes #10746

This PR adds a new introductory vignette which replaces most of the previous ones, and modifies the code examples throughout functions aimed at interactive usage to call xgboost() instead of xgb.train().

Motivation

Since the time that XGBoost was first published at CRAN, its adoption and mindshare have risen substantially, to the point that it has become the standard when it comes to boosted decision trees. In this day and age, I don't think the package needs to provide any introduction to the concepts of gradient-boosting, cross-validation, evaluation metrics, and so on - people who use R are already going to be familiar with those, and the things it compares against (like the package 'gbm') have become obsolete by now.

As well, the documentation and tutorials for XGBoost have mostly moved to the online docs - any R-specific documents become outdated rather soon, and are less likely to be seen by a random user. Most of the python examples and guides should in any event work with the R interface with very minimal modifications like dict->list.

Apart from becoming a standard-use library, the features supported by XGBoost have expanded over time, and lots of the materials that were there before, such as the first vignette, contained tips that are not applicable to the current state of the library, like manually one-hot encoding categorical features.

Hence, I decided to remove the previous vignettes and create a new one from scratch, which contains only examples around the usage of the R interface and its conventions.

Help needed

It would be ideal if this vignette could also get added to the online docs.

Thus, I created the vignette as a quarto file (.qmd), which has the option to render to both .html (what CRAN hosts) and .md (which can be included in the .rst files).

Only, getting it to render to .md required building the vignette with jupyter instead of knitr, which in turn requires installs of python, jupyter, ipykernel, and the "ir" kernel that runs R in jupyter, plus registering that kernel in the user-level config for jupyter. By adding that line "jupyter: ir", it additionally makes the default quarto render (e.g. as used by the "knit" button in RStudio) build the .html vignette using jupyter instead of knitr, which is most definitely not going to work out in CRAN servers. I don't know how to solve this.

Would also be nice if some CI job could be auto-building the .md file for the online docs from the .qmd source of the vignette.

(CCing @mayer79 and @jameslamb )

[jameslamb]

Thanks for the @. I haven't reviewed the content of this PR closely so am limiting my comment only to the topic of the "Help Needed" part above.

I know nothing about quarto or how to integrate it with CRAN's vignette-building and vignette-checking. Maybe these examples of other packages using it successfully on CRAN could lead to some good reference implementations: https://github.com/search?q=org%3Acran%20path%3A*.qmd&type=code

I can at least tell you how we're addressing these things LightGBM, maybe it will be helpful.

In {lightgbm}, we have a lightweight system that has been working well for us (which you alluded to in #9906 (comment)). Here's how we do this there:

1. vignettes are written in .Rmd files with headmatter like the following, using {markdown} + {knitr}

output:
  markdown::html_format:
    options:
      toc: true
      number_sections: true
vignette: >
  %\VignetteIndexEntry{Basic Walkthrough}
  %\VignetteEngine{knitr::knitr}
  %\VignetteEncoding{UTF-8}

(code link)

2. The main docs site is built via Sphinx, and the Sphinx build builds a {pkgdown} site with the roxygen docs + vignettes + demos (code link)
3. those docs are integrated into the readthedocs site using an absolute hyperlink (code link)
4. these are linked from the "R API" nav item at https://lightgbm.readthedocs.io/en/latest/index.html, which directs to https://lightgbm.readthedocs.io/en/latest/R/reference/

It's a little awkward that the R docs are not technically part of the readthedocs site, but overall this has worked pretty well for us. Some notes:

• {markdown} + {knitr} is a lighter-weight combination than using {rmarkdown} or {quarto} (fewer required dependencies to install)
  • ref: [R-package] [ci] switch vignettes from 'rmarkdown' to 'markdown' microsoft/LightGBM#6258
• there are many examples of using .Rmd files + this combination on CRAN, so staying CRAN-compliant requires very little ongoing maintenance
• {pkgdown} is customizable enough to meet our needs, and it keeps up with changes to the RMarkdown format

Even if you don't pursue this specific mix, I do recommend not checking the .md version of the vignette into source control here. Instead, you could put any commands to generate it as part of the docs-site building into doc/conf.py or doc/Makefile. I shared a link above to how to run custom commands at build-time via conf.py. Here's an example of how to do it via the Makefile generate by sphinx: https://github.com/rapidsai/legate-boost/blob/ba062bf666845f82f5e8ab4690c97db4469ab42f/docs/Makefile#L17-L29