def test_bgl_unfair(A_two_dim):
a0_count = 5
a1_count = 7
a0_label = 2
a1_label = 3
a0_factor = 1
a1_factor = 16
grid_size = 7
X, Y, A = _simple_regression_data(
a0_count, a1_count, a0_factor, a1_factor, a0_label, a1_label, A_two_dim
)
bgl_square_loss = BoundedGroupLoss(SquareLoss(-np.inf, np.inf))
grid_search = GridSearch(
LinearRegression(), constraints=bgl_square_loss, grid_size=grid_size
)
grid_search.fit(X, Y, sensitive_features=A)
assert_n_grid_search_results(grid_size, grid_search)
test_X = pd.DataFrame(
{
"actual_feature": [0.2, 0.7],
"sensitive_features": [a0_label, a1_label],
"constant_ones_feature": [1, 1],
}
)
best_predict = grid_search.predict(test_X)
assert np.allclose([-1.91764706, 9.61176471], best_predict)
all_predict = [predictor.predict(test_X) for predictor in grid_search.predictors_]
# TODO: investigate where the different outcomes for the first grid point are from, likely
# due to some ignored data points at the edge resulting in another solution with the same
# least squares loss (i.e. both solutions acceptable).
# Reflects https://github.com/fairlearn/fairlearn/issues/265
assert logging_all_close([[3.2, 11.2]], [all_predict[0]]) or logging_all_close(
[[3.03010885, 11.2]], [all_predict[0]]
)
assert logging_all_close(
[
[-3.47346939, 10.64897959],
[-2.68, 10.12],
[-1.91764706, 9.61176471],
[-1.18461538, 9.12307692],
[-0.47924528, 8.65283019],
[0.2, 0.7],
],
all_predict[1:],
)
def test_bgl_unfair(A_two_dim):
a0_count = 5
a1_count = 7
a0_label = 2
a1_label = 3
a0_factor = 1
a1_factor = 16
X, Y, A = _simple_regression_data(a0_count, a1_count, a0_factor, a1_factor,
a0_label, a1_label, A_two_dim)
bgl_square_loss = GroupLossMoment(SquareLoss(-np.inf, np.inf))
target = GridSearch(LinearRegression(),
constraints=bgl_square_loss,
grid_size=7)
target.fit(X, Y, sensitive_features=A)
assert len(target.all_results) == 7
test_X = pd.DataFrame({
"actual_feature": [0.2, 0.7],
"sensitive_features": [a0_label, a1_label],
"constant_ones_feature": [1, 1]
})
best_predict = target.predict(test_X)
assert np.allclose([-1.91764706, 9.61176471], best_predict)
all_predict = [r.predictor.predict(test_X) for r in target.all_results]
assert logging_all_close(
[[3.2, 11.2], [-3.47346939, 10.64897959], [-2.68, 10.12],
[-1.91764706, 9.61176471], [-1.18461538, 9.12307692],
[-0.47924528, 8.65283019], [0.2, 0.7]], all_predict)