FIX ignore nan values when summing posteriors

[PSSF23]

Close #290

• Fix np.nan posteriors when honest_prior="ignore"
• Set "ignore" as default prior
• Add & optimize unit tests

[PSSF23]

Seems a simple and straightforward fix to me. I'll write unit tests about it, but first want to see if it cause other problems.

[PSSF23]

This test is not checking the ignore setting properly: test_impute_posteriors. As it doesn't give an error when there's too many nan's.

The fix is also incomplete, as np.nansum gives zeros when all the tree posteriors are nan's. impute_missing is not working as it should. I'm forcing it to work on zero mask posteriors, as the other two leaf correction methods would not result in zeros.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 78.53%. Comparing base (574a629) to head (c4573ae).
Report is 4 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #291      +/-   ##
==========================================
- Coverage   78.55%   78.53%   -0.02%     
==========================================
  Files          24       24              
  Lines        2252     2250       -2     
  Branches      414      413       -1     
==========================================
- Hits         1769     1767       -2     
  Misses        352      352              
  Partials      131      131              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[adam2392]

Left a few comments.

Re CI failures:

FAILED sktree/stats/tests/test_forestht.py::test_small_dataset_dependent[0] - AssertionError
FAILED sktree/tests/test_extensions.py::test_predict_proba_per_tree[HonestForestClassifier-2] - AssertionError
FAILED sktree/tests/test_extensions.py::test_predict_proba_per_tree[HonestForestClassifier-3] - AssertionError

I would recommend:

1. test_small_dataset_dependent, just adding the honest_prior='empirical' to make the test backwards compatible.
2. test_predict_proba_per_tree, add an extra kwarg for HonestForestClassifier for the honest_prior='empirical'. You can also add a test for test_predict_proba_per_tree for HonestForestClassifier if honest_prior='ignore'.

[PSSF23]

@adam2392 you think this is mergeable?

[adam2392]

Thanks for the PR @PSSF23