Typo in grid_search_reduction.py causing AttributeError: 'GridSearchReduction' object has no attribute 'moment'

[haas-christian]

In the current version of GridSearchReduction, there's a typo in the predict() function that causes an AttributeError when trying to use a fitted GridSearchReduction model for predicting a new dataset.

Specifically, the typo is in line 140:
if isinstance(self.model.moment, red.ClassificationMoment):

Since the model does not have an attribute moment, it will cause the Attribute Error. Instead, the moment attribute of the model should be self.model.moment_ :
if isinstance(self.model.moment_, red.ClassificationMoment):

Here's an example that shows the described behavior:

# Load all necessary packages
from aif360.algorithms.preprocessing.optim_preproc_helpers.data_preproc_functions import load_preproc_data_adult
from aif360.algorithms.inprocessing.grid_search_reduction import GridSearchReduction
from sklearn.linear_model import LogisticRegression
import numpy as np

# Get the dataset and split into train and test
dataset_orig = load_preproc_data_adult()

privileged_groups = [{'sex': 1}]
unprivileged_groups = [{'sex': 0}]

np.random.seed(0)
dataset_orig_train, dataset_orig_test = dataset_orig.split([0.7], shuffle=True)

estimator = LogisticRegression(solver='lbfgs', max_iter=1000)

np.random.seed(0) #need for reproducibility
grid_red = GridSearchReduction(estimator=estimator,
                               constraints="EqualizedOdds",
                               drop_prot_attr=False)
grid_red.fit(dataset_orig_train)
grid_red_pred = grid_red.predict(dataset_orig_test)

[mnagired]

Hi there,

Thanks for flagging this. Interesting, do you know if self.model.moment_ is the only way to fix this?

For example, in the code for GridSearchReduction in the sklearn-compatible version of AIF360, you can see the use of self.model._constraints.

[haas-christian]

That's a good question. Based on my understanding, we want to check if the provided moment is one that is implemented as ClassificationMoment in fairlearn. if isinstance(self.model.moment_, red.ClassificationMoment) would check this, since self.moment_ in the sklearn-compatible version of GridSearchReduction refers to the fairlearn ClassificationMoment. In contrast, if we call self.model.constraints, we only get the str representation.

In the above code sample, when I call self.model.moment_, I get the result <fairlearn.reductions._moments.utility_parity.EqualizedOdds at 0x1bb41350130>, indicating it's a FairLearn moment.
However, if I call self.model.constraints, I get 'EqualizedOdds', indicating a simple str representation.

An alternative way to checking if isinstance(self.model.moment, red.ClassificationMoment) could be to check if self.model.constraints is a key in a dict of accepted moments, similar to what the sklearn-compatible version does (lines 116-123):

        moments = {
            "DemographicParity": red.DemographicParity,
            "EqualizedOdds": red.EqualizedOdds,
            "TruePositiveRateParity": red.TruePositiveRateParity,
            "FalsePositiveRateParity": red.FalsePositiveRateParity,
            "ErrorRateParity": red.ErrorRateParity,
            "BoundedGroupLoss": red.BoundedGroupLoss,
        }

However, since this dict is only specified in fit(), it cannot be accessed via self.model at the moment. So, another fix could be to add this dict of accepted moments in the init() function of GridSearchReduction, and then check if the model moment is in this dict, similar to what lines 124-126 do in the sklearn-compatible version:

if isinstance(self.constraints, str):
            if self.constraints not in moments:
                raise ValueError(f"Constraint not recognized: {self.constraints}")

Compared to this, fixing the typo to self.model.moment_ seems the easier choice in my opinion.

[mnagired]

Thanks for the exploration!

This makes sense to me, could you please open a PR with your changes to line 140 above and link this issue there?

[haas-christian]

Done!