[R-package] ensure use of interaction_constraints does not lead to features being ignored

[mayer79]

This enhances the R-API of interaction constraints by adding a feature group with those features that do not appear in any of the interaction groups. Currently, these are simply dropped from training, which seems undesirable.

Additionally, it reorganizes the code of the corresponding helper function .check_interaction_constraints().

It solves the R-package part of #6376. I will attempt a separate PR for the Python-package.

Example

.check_interaction_constraints(list("a", 2:3), letters[1:5])

# Output
[[1]]
[1] "[0]"

[[2]]
[1] "[1,2]"

[[3]]
[1] "[3,4]"

Without the PR, the result is

[[1]]
[1] "[0]"

[[2]]
[1] "[1,2]"

i.e., the last two features are silently dropped from the training.

[mayer79]

Getting a failing unit test:

  -- 1. Failure (test_lgb.Booster.R:209:5): Loading a Booster from a text file wor

  bst2$params[names(params)] not equal to `params`.

  Component "interaction_constraints": Length mismatch: comparison on first 2 components

I will have a look at it next week (afk).

[jameslamb]

Thanks! Left a few minor comments.

@ me if you need help with the failing tests. Let's also please get confirmation that excluding omitted features from training wasn't intentional (#6376 (comment)).

[jameslamb]

Hey @mayer79 , across your recent PRs I've seen multiple "fix linting" types of commits. Totally fine to keep using Continuous Integration to get that feedback (we don't have a lot of activity going on in the repo right now), but you'd probably find it faster to run the linting locally. It only requires R and the {lintr} package.

Rscript .ci/lint_r_code.R $(pwd)/R-package

[mayer79]

Hey @mayer79 , across your recent PRs I've seen multiple "fix linting" types of commits. Totally fine to keep using Continuous Integration to get that feedback (we don't have a lot of activity going on in the repo right now), but you'd probably find it faster to run the linting locally. It only requires R and the {lintr} package.

Rscript .ci/lint_r_code.R $(pwd)/R-package

Thanks, this is the stuff that I should have asked long time ago, but never did :-).

[mayer79]

Pipeline seems happy @jameslamb - but really no pressure :-)

[jameslamb]

Thanks very much for this! And I really appreciate you keeping it up to date with master while waiting for reviews.

I just left one suggestion about covering this a bit more thoroughly with tests. If you don't have time in the next week to get to that, let me know if you're open to me writing that test and pushing it here...I'm sorry for the time pressure (especially after not reviewing this for so long 😬 ), but I'm going to try to push for v4.4.0 to get out in the next week, ahead of the numpy 2.0 release (#6439 (comment)), and I'd love to get this change into it if we can.

[mayer79]

Thanks very much for this! And I really appreciate you keeping it up to date with master while waiting for reviews.

I just left one suggestion about covering this a bit more thoroughly with tests. If you don't have time in the next week to get to that, let me know if you're open to me writing that test and pushing it here...I'm sorry for the time pressure (especially after not reviewing this for so long 😬 ), but I'm going to try to push for v4.4.0 to get out in the next week, ahead of the numpy 2.0 release (#6439 (comment)), and I'd love to get this change into it if we can.

Thanks so much for the review! I will try to improve the tests this week.

[jameslamb]

Changes look great to me, thanks very much @mayer79 !

This will get into v4.4.0 😁

[jameslamb]

It seems that CI here is suffering from this issue: tox-dev/filelock#337

(build link)

That issue was just discovered earlier today, and fixes for it are still rolling out. To keep making progress towards the release, I've just temporarily commented out running pre-commit in CI. I strongly suspect that 24 hours from now, that'll be resolved and we can re-enable it.

[github-actions]

This pull request has been automatically locked since there has not been any recent activity since it was closed.
To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this.