Robust computation of the probability weights

[NglQ]

Robust probability computation of weights

Traceback (most recent call last):
  File "postprocessing.py", line 240, in <module>
    pipeline(load_fn=dataset_name_to_load_fn[dataset_name],
  File "postprocessing.py", line 208, in pipeline
    unprocessed_clf.fit(X=data_splits['X_' + fit_data], y=data_splits['y_' + fit_data], group=data_splits['s_' + fit_data])
  File "error_parity/threshold_optimizer.py", line 485, in fit
    g: RandomizedClassifier.construct_at_target_ROC(
       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "error_parity/classifiers.py", line 372, in construct_at_target_ROC
    weights, points = RandomizedClassifier.find_weights_given_two_points(
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "error_parity/classifiers.py", line 283, in find_weights_given_two_points
    raise RuntimeError(
RuntimeError: Triangulation of target point failed. Target was [5.65363161e-07 5.42325664e-07]; got [5.42325665e-07 5.42325665e-07].

The issue has been found using "ACSDataSource(survey_year='2018', horizon='1-Year', survey='person')" as dataset, with the basic problem "ACSEmployment". The split is 70-10-20 using the function train_test_split of sklearn with shuffle=True, stratify on sensitive attribute (RAC1P) and seed 42. Moreover, the RelaxedThreshold had equalized odds as constraint and threshold 1.0 (The experiment involves FairGBM as model).

[AndreFCruz]

Thanks for reporting this bug!
I'll double-check everything is fine and accept the atol modification in the next few days 🙂

[AndreFCruz]

Did you ever encounter a scenario where any weights were negative? And if so, could you post a code example?