def _selected_label_compare(moment, metric, selected_label):
# Similar to _simple_compare, but we need to worry about the y label
X, y = loan_scenario_generator(n, f, sfs, ibs, seed=7132752)
X_dummy = pd.get_dummies(X)
est = LogisticRegression()
est.fit(X_dummy, y)
y_pred = est.predict(X_dummy)
target = moment()
target.load_data(np.asarray(X_dummy), np.asarray(y),
sensitive_features=X['sens'],
control_features=X['ctrl'])
# gamma measures the constraint violation relative to the overall value
results = target.gamma(est.predict)
# Compute the constraint violation using the metrics
mf_pred = MetricFrame(metrics=metric, y_true=y, y_pred=y_pred,
sensitive_features=X['sens'],
control_features=X['ctrl'])
diffs = mf_pred.by_group - mf_pred.overall
# Compare (with a very small amount of wriggle room)
for ib in ibs:
for sf in sfs:
# Format defined within utility_parity._combine_event_and_control
label_format = "control={0},label={1}"
label = label_format.format(ib, selected_label)
assert diffs[(ib, sf)] == pytest.approx(results[('+', label, sf)],
rel=1e-10, abs=1e-12)
assert diffs[(ib, sf)] == pytest.approx(-results[('-', label, sf)],
rel=1e-10, abs=1e-12)
def _simple_compare(moment, metric):
X, y = loan_scenario_generator(n, f, sfs, ibs, seed=7632752)
X_dummy = pd.get_dummies(X)
est = LogisticRegression()
est.fit(X_dummy, y)
y_pred = est.predict(X_dummy)
target = moment()
target.load_data(np.asarray(X_dummy), np.asarray(y),
sensitive_features=X['sens'],
control_features=X['ctrl'])
# gamma measures the constraint violation relative to the overall value
results = target.gamma(est.predict)
# Compute the constraint violation using the metrics
mf_pred = MetricFrame(metrics=metric, y_true=y, y_pred=y_pred,
sensitive_features=X['sens'],
control_features=X['ctrl'])
diffs = mf_pred.by_group - mf_pred.overall
# Compare (with a very small amount of wriggle room)
for ib in ibs:
for sf in sfs:
event_format = "control={0},all"
assert diffs[(ib, sf)] == pytest.approx(results[('+', event_format.format(ib), sf)],
rel=1e-10, abs=1e-12)
assert diffs[(ib, sf)] == pytest.approx(-results[('-', event_format.format(ib), sf)],
rel=1e-10, abs=1e-12)
def test_equalized_odds():
# Have to do this one longhand, since it combines tpr and fpr
X, y = loan_scenario_generator(n, f, sfs, ibs, seed=632753)
X_dummy = pd.get_dummies(X)
metrics = {"tpr": true_positive_rate, "fpr": false_positive_rate}
unmitigated = LogisticRegression()
unmitigated.fit(X_dummy, y)
y_pred = unmitigated.predict(X_dummy)
mf_unmitigated = MetricFrame(
metrics=metrics,
y_true=y,
y_pred=y_pred,
sensitive_features=X["sens"],
control_features=X["ctrl"],
)
expgrad_basic = ExponentiatedGradient(
LogisticRegression(),
constraints=EqualizedOdds(difference_bound=0.01),
eps=0.01)
expgrad_basic.fit(X_dummy, y, sensitive_features=X["sens"])
y_pred_basic = expgrad_basic.predict(X_dummy, random_state=9235)
mf_basic = MetricFrame(
metrics=metrics,
y_true=y,
y_pred=y_pred_basic,
sensitive_features=X["sens"],
control_features=X["ctrl"],
)
expgrad_control = ExponentiatedGradient(
LogisticRegression(),
constraints=EqualizedOdds(difference_bound=0.01),
eps=0.01)
expgrad_control.fit(X_dummy,
y,
sensitive_features=X["sens"],
control_features=X["ctrl"])
y_pred_control = expgrad_control.predict(X_dummy, random_state=8152)
mf_control = MetricFrame(
metrics=metrics,
y_true=y,
y_pred=y_pred_control,
sensitive_features=X["sens"],
control_features=X["ctrl"],
)
compare_unmitigated = mf_control.difference(
method="to_overall") <= mf_unmitigated.difference(method="to_overall")
print(compare_unmitigated)
compare_basic = mf_control.difference(
method="to_overall") <= mf_basic.difference(method="to_overall")
print(compare_basic)
assert compare_basic.values.reshape(6).all()
assert compare_unmitigated.values.reshape(6).all()
def run_comparisons(moment, metric_fn):
X, y = loan_scenario_generator(n, f, sfs, ibs, seed=163)
X_dummy = pd.get_dummies(X)
mf_input = MetricFrame(metric_fn, y, y,
sensitive_features=X['sens'],
control_features=X['ctrl'])
print("Metric for input:\n", mf_input.by_group)
print("Input Metric differences:\n", mf_input.difference(method='to_overall'), "\n")
unmitigated = LogisticRegression()
unmitigated.fit(X_dummy, y)
y_pred = unmitigated.predict(X_dummy)
mf_unmitigated = MetricFrame(metric_fn,
y, y_pred,
sensitive_features=X['sens'],
control_features=X['ctrl'])
print("Unmitigated metric:\n", mf_unmitigated.by_group)
print("Unmitigated metric differences:\n",
mf_unmitigated.difference(method='to_overall'), "\n")
expgrad_basic = ExponentiatedGradient(
LogisticRegression(),
constraints=moment(),
eps=0.005)
expgrad_basic.fit(X_dummy, y, sensitive_features=X['sens'])
y_pred_basic = expgrad_basic.predict(X_dummy, random_state=8235)
mf_basic = MetricFrame(metric_fn, y, y_pred_basic,
sensitive_features=X['sens'],
control_features=X['ctrl'])
print("Basic expgrad metric:\n", mf_basic.by_group)
print("Basic expgrad metric differences:\n",
mf_basic.difference(method='to_overall'), "\n")
expgrad_control = ExponentiatedGradient(
LogisticRegression(),
constraints=moment(),
eps=0.005)
expgrad_control.fit(X_dummy, y,
sensitive_features=X['sens'],
control_features=X['ctrl'])
y_pred_control = expgrad_control.predict(X_dummy, random_state=852)
mf_control = MetricFrame(metric_fn, y, y_pred_control,
sensitive_features=X['sens'],
control_features=X['ctrl'])
print("expgrad_control metric:\n", mf_control.by_group)
print("expgrad_control metric differences:\n",
mf_control.difference(method='to_overall'))
assert (mf_control.difference(method='to_overall') <=
mf_unmitigated.difference(method='to_overall')).all()
assert (mf_control.difference(method='to_overall') <=
mf_basic.difference(method='to_overall')).all()
