def audit(df, configs, model_id=1, preprocessed=False):
"""
:param df:
:param ref_groups_method:
:param model_id:
:param configs:
:param report:
:param preprocessed:
:return:
"""
if not preprocessed:
df, attr_cols_input = preprocess_input_df(df)
if not configs.attr_cols:
configs.attr_cols = attr_cols_input
g = Group()
print('Welcome to Aequitas-Audit')
print('Fairness measures requested:',
','.join(configs.fair_measures_requested))
groups_model, attr_cols = g.get_crosstabs(
df,
score_thresholds=configs.score_thresholds,
model_id=model_id,
attr_cols=configs.attr_cols)
print('audit: df shape from the crosstabs:', groups_model.shape)
b = Bias()
# todo move this to the new configs object / the attr_cols now are passed through the configs object...
ref_groups_method = configs.ref_groups_method
if ref_groups_method == 'predefined' and configs.ref_groups:
bias_df = b.get_disparity_predefined_groups(groups_model, df,
configs.ref_groups)
elif ref_groups_method == 'majority':
bias_df = b.get_disparity_major_group(groups_model, df)
else:
bias_df = b.get_disparity_min_metric(groups_model, df)
print('Any NaN?: ', bias_df.isnull().values.any())
print('bias_df shape:', bias_df.shape)
aqp = Plot()
if len(configs.plot_bias_metrics) == 1:
fig1 = aqp.plot_disparity(bias_df, metrics=configs.plot_bias_metrics)
elif len(configs.plot_bias_metrics) > 1:
fig1 = aqp.plot_disparity_all(bias_df,
metrics=configs.plot_bias_metrics)
if len(configs.plot_bias_disparities) == 1:
fig2 = aqp.plot_group_metric(bias_df,
metrics=configs.plot_bias_disparities)
elif len(configs.plot_bias_disparities) > 1:
fig2 = aqp.plot_group_metric_all(bias_df,
metrics=configs.plot_bias_disparities)
f = Fairness(tau=configs.fairness_threshold)
print('Fairness Threshold:', configs.fairness_threshold)
print('Fairness Measures:', configs.fair_measures_requested)
group_value_df = f.get_group_value_fairness(
bias_df, fair_measures_requested=configs.fair_measures_requested)
group_attribute_df = f.get_group_attribute_fairness(
group_value_df,
fair_measures_requested=configs.fair_measures_requested)
fair_results = f.get_overall_fairness(group_attribute_df)
if len(configs.plot_bias_metrics) == 1:
fig3 = aqp.plot_fairness_group(group_value_df,
metrics=configs.plot_bias_metrics)
elif len(configs.plot_bias_metrics) > 1:
fig3 = aqp.plot_fairness_group_all(group_value_df,
metrics=configs.plot_bias_metrics)
if len(configs.plot_bias_disparities) == 1:
fig4 = aqp.plot_fairness_disparity(
group_value_df, metrics=configs.plot_bias_disparities)
elif len(configs.plot_bias_metrics) > 1:
fig4 = aqp.plot_fairness_disparity_all(
group_value_df, metrics=configs.plot_bias_disparities)
print(fair_results)
report = None
if configs.report is True:
report = audit_report_markdown(configs, group_value_df,
f.fair_measures_depend, fair_results)
return group_value_df, report
hbdf = b.get_disparity_predefined_groups(xtab,
original_df=df,
ref_groups_dict={
'race': 'WHITE',
'gender': 'MALE'
},
alpha=0.5,
mask_significance=False)
majority_bdf = b.get_disparity_major_group(xtab,
original_df=df,
mask_significance=False)
majority_bdf[['attribute_name', 'attribute_value'] + calculated_disparities +
disparity_significance]
tm_capped = aqp.plot_disparity_all(hbdf,
attributes=['gender', 'race'],
metrics='all',
significance_alpha=0.05)
f = Fairness()
fdf = f.get_group_value_fairness(bdf)
parity_detrminations = f.list_parities(fdf)
fdf[['attribute_name', 'attribute_value'] + absolute_metrics +
calculated_disparities + parity_detrminations].style
fg = aqp.plot_fairness_group_all(fdf, ncols=5, metrics="all")
n_tm = aqp.plot_fairness_disparity_all(fdf,
attributes=['race', 'gender'],
significance_alpha=0.05)