[ENH] Skeleton learning method

[adam2392]

Follow on to: #19
And part of #18

Changes proposed in this pull request:

Before submitting

• I've read and followed all steps in the Making a pull request
  section of the CONTRIBUTING docs.
• I've updated or added any relevant docstrings following the syntax described in the
  Writing docstrings section of the CONTRIBUTING docs.
• If this PR fixes a bug, I've added a test that will fail without my fix.
• If this PR adds a new feature, I've added tests that sufficiently cover my new functionality.

After submitting

• All GitHub Actions jobs for my pull request have passed.

[codecov-commenter]

Codecov Report

Merging #20 (ead714f) into main (b56a106) will decrease coverage by 1.62%.
The diff coverage is 68.32%.

@@            Coverage Diff             @@
##             main      #20      +/-   ##
==========================================
- Coverage   76.83%   75.21%   -1.63%     
==========================================
  Files           8       11       +3     
  Lines         367      585     +218     
  Branches       59      116      +57     
==========================================
+ Hits          282      440     +158     
- Misses         66      104      +38     
- Partials       19       41      +22     

Impacted Files	Coverage Δ
dodiscover/constraint/utils.py	46.15% <46.15%> (ø)
dodiscover/ci/oracle.py	86.95% <50.00%> (+30.43%)	⬆️
dodiscover/context.py	67.39% <67.39%> (ø)
dodiscover/constraint/skeleton.py	70.51% <70.51%> (ø)
dodiscover/_protocol.py	61.53% <75.00%> (+0.66%)	⬆️
dodiscover/ci/g_test.py	63.63% <0.00%> (+1.65%)	⬆️

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[adam2392]

This now gives us a generic backbone for all the PC/FCI algo variants, where the first step is "learning a skeleton" from CI testing of the data.

The LearnSkeleton can most likely be subclassed for additional structure such as computing the "possibly d-separating sets" in RFCI/FCI and evaluation of F-nodes when given interventional data.

[bloebp]

Hmm, 'tuning' the parameter in a way to make the test work is maybe not a good idea. There might be something wrong with the implementation, which is then not visible due to the tuned parameter. If this is purely due to the flakiness of the (statistical) tests, consider rerunning them them when they fail with newly generated data. If it fails, e.g., 3x in a row, then it is a good indication to double check the implementation (and this could be the independence test here). See for instance https://github.com/py-why/dowhy/blob/main/tests/gcm/independence_test/test_kernel.py#L12, where we use flaky for exactly these kind of flaky scenarios.

[adam2392]

IIRC I think I took the data from pcalg of the R package... so it's not actually generated on the fly.

I assumed it was fine since the tests passed other types of simulated data, but let me actually investigate.

I might just play around w/ simulating some new data instead from an SCM.

[adam2392]

Oh wait I just checked again and 0.05 works. I think this might've been a different private project that I was working on. The alpha level works for 0.05 (default) and 0.01.