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,
configs.ref_groups)
elif ref_groups_method == 'majority':
bias_df = b.get_disparity_major_group(groups_model)
else:
bias_df = b.get_disparity_min_metric(groups_model)
print('Any NaN?: ', bias_df.isnull().values.any())
print('bias_df shape:', bias_df.shape)
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)
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
def get_model_fairness(self, level='model'):
g = Group()
xtab, _ = g.get_crosstabs(self.df)
b = Bias()
majority_bdf = b.get_disparity_major_group(xtab,
original_df=self.df,
mask_significance=True)
f = Fairness()
fdf = f.get_group_value_fairness(majority_bdf)
f_res = fdf
if level == 'model':
f_res = f.get_overall_fairness(fdf)
elif level == 'attribute':
f_res = f.get_group_attribute_fairness(fdf)
return f_res
def fairness(df):
"""
Genera todo el módulo de Equidad
"""
print("Módulo de Equidad")
print("-"*30)
f = Fairness()
bias_ = Bias()
g = Group()
xtab, atts = g.get_crosstabs(df, attr_cols=["delegacion"])
absolute_metrics = g.list_absolute_metrics(xtab)
bdf = bias_.get_disparity_predefined_groups(xtab, original_df = df,
ref_groups_dict = {'delegacion': 'IZTAPALAPA'},
alpha=0.05)
fdf = f.get_group_value_fairness(bdf)
parity_determinations = f.list_parities(fdf)
print("Imprimiendo tabla de métricas (conteos en frecuencias):")
print(fdf[['attribute_name', 'attribute_value'] + absolute_metrics + bias_.list_disparities(fdf) + parity_determinations].round(2))
print("Impriendo métricas generales")
gof = f.get_overall_fairness(fdf)
print(gof)
print("Aequitas analysis completed.")
print("-"*30)
def fairnessf(bdf, absolute_metrics, bias):
"""
args:
df (dataframe): Recibe el data frame que tiene los features sobre los que queremos medir la equidad.
returns:
-
"""
fair = Fairness()
fdf = fair.get_group_value_fairness(bdf)
parity_determinations = fair.list_parities(fdf)
fairness = fdf[['attribute_name', 'attribute_value'] + absolute_metrics +
bias.list_disparities(fdf) + parity_determinations].round(2)
## Storing metadata
aq_metadata["v_group"] = str(fdf.loc[0, "attribute_value"])
aq_metadata["FOR_p"] = str(fdf.loc[0, "FOR Parity"])
aq_metadata["FNR_p"] = str(fdf.loc[0, "FNR Parity"])
#return df_aeq
gaf = fair.get_group_attribute_fairness(fdf)
gof = fair.get_overall_fairness(fdf)
return fairness, gaf, gof
def run_aequitas(predictions_data_path):
'''
Check for False negative rate, chances of certain group missing out on assistance using aequitas toolkit
The functions transform the data to make it aequitas complaint and checks for series of bias and fairness metrics
Input: model prediction path for the selected model (unzip the selected file to run)
Output: plots saved in charts folder
'''
best_model_pred = pd.read_csv(predictions_data_path)
# Transform data for aquetias module compliance
aqc = [
'Other', 'White', 'African American', 'Asian', 'Hispanic',
'American Indian'
]
aqcol = [
'White alone_scale', 'Black/AfAmer alone_scale',
'AmInd/Alaskn alone_scale', 'Asian alone_scale', 'HI alone_scale',
'Some other race alone_scale', 'Hispanic or Latino_scale'
]
display(aqcol)
aqcol_label = [
'no_renew_nextpd', 'pred_class_10%',
'Median household income (1999 dollars)_scale'
] + aqcol
aqus = best_model_pred[aqcol_label]
print('Creating classes for racial and income distribution', '\n')
# Convert to binary
bin_var = [
'no_renew_nextpd',
'pred_class_10%',
]
for var in bin_var:
aqus[var] = np.where(aqus[var] == True, 1, 0)
# Rename
aqus.rename(columns={
'no_renew_nextpd': 'label_value',
'pred_class_10%': 'score'
},
inplace=True)
print('Define majority rule defined on relative proportion of the class',
'\n')
aqus['race'] = aqus[aqcol].idxmax(axis=1)
# Use quantile income distribution
aqus['income'] = pd.qcut(
aqus['Median household income (1999 dollars)_scale'],
3,
labels=["rich", "median", "poor"])
# Final form
aqus.drop(aqcol, axis=1, inplace=True)
aqus.drop(['Median household income (1999 dollars)_scale'],
axis=1,
inplace=True)
aq = aqus.reset_index()
aq.rename(columns={'index': 'entity_id'}, inplace=True)
aq['race'] = aq['race'].replace({
'Some other race alone_scale':
'Other',
'White alone_scale':
'White',
'Black/AfAmer alone_scale':
'African American',
'Asian alone_scale':
'Asian',
'HI alone_scale':
'Hispanic',
'AmInd/Alaskn alone_scale':
'American Indian'
})
# Consolidate types
aq['income'] = aq['income'].astype(object)
aq['entity_id'] = aq['entity_id'].astype(object)
aq['score'] = aq['score'].astype(object)
aq['label_value'] = aq['label_value'].astype(object)
# Distribuion of categories
aq_palette = sns.diverging_palette(225, 35, n=2)
by_race = sns.countplot(x="race", data=aq[aq.race.isin(aqc)])
by_race.set_xticklabels(by_race.get_xticklabels(), rotation=40, ha="right")
plt.savefig('charts/Racial distribution in data.png')
# Primary distribuion against score
aq_palette = sns.diverging_palette(225, 35, n=2)
by_race = sns.countplot(x="race",
hue="score",
data=aq[aq.race.isin(aqc)],
palette=aq_palette)
by_race.set_xticklabels(by_race.get_xticklabels(), rotation=40, ha="right")
# Race
plt.savefig('charts/race_score.png')
# Income
by_inc = sns.countplot(x="income",
hue="score",
data=aq,
palette=aq_palette)
plt.savefig('charts/income_score.png')
# Set Group
g = Group()
xtab, _ = g.get_crosstabs(aq)
# False Negative Rates
aqp = Plot()
fnr = aqp.plot_group_metric(xtab, 'fnr', min_group_size=0.05)
p = aqp.plot_group_metric_all(xtab,
metrics=['ppr', 'pprev', 'fnr', 'fpr'],
ncols=4)
p.savefig('charts/eth_metrics.png')
# Bias with respect to white rich category
b = Bias()
bdf = b.get_disparity_predefined_groups(xtab,
original_df=aq,
ref_groups_dict={
'race': 'White',
'income': 'rich'
},
alpha=0.05,
mask_significance=True)
bdf.style
calculated_disparities = b.list_disparities(bdf)
disparity_significance = b.list_significance(bdf)
aqp.plot_disparity(bdf,
group_metric='fpr_disparity',
attribute_name='race',
significance_alpha=0.05)
plt.savefig('charts/disparity.png')
# Fairness
hbdf = b.get_disparity_predefined_groups(xtab,
original_df=aq,
ref_groups_dict={
'race': 'African American',
'income': 'poor'
},
alpha=0.05,
mask_significance=False)
majority_bdf = b.get_disparity_major_group(xtab,
original_df=aq,
mask_significance=True)
min_metric_bdf = b.get_disparity_min_metric(df=xtab, original_df=aq)
f = Fairness()
fdf = f.get_group_value_fairness(bdf)
parity_detrminations = f.list_parities(fdf)
gaf = f.get_group_attribute_fairness(fdf)
gof = f.get_overall_fairness(fdf)
z = aqp.plot_fairness_group(fdf, group_metric='ppr')
plt.savefig('charts/fairness_overall.png')
# Checking for False Omission Rate and False Negative Rates
fg = aqp.plot_fairness_group_all(fdf, metrics=['for', 'fnr'], ncols=2)
fg.savefig('charts/fairness_metrics.png')
return None
def fun_bias_fair(a_zip, a_type, fea_eng, model):
X = fea_eng.drop([
'aka_name', 'facility_type', 'address', 'inspection_date',
'inspection_type', 'violations', 'results', 'pass'
],
axis=1)
y_pred = model.predict(X)
xt = pd.DataFrame([
fea_eng['zip'].astype(float), fea_eng['facility_type'],
fea_eng['pass'], y_pred
]).transpose()
a_zip['zip'] = a_zip['zip'].astype(float)
compas = pd.merge(left=xt,
right=a_zip,
how='left',
left_on='zip',
right_on='zip')
compas = pd.merge(left=compas,
right=a_type,
how='left',
left_on='facility_type',
right_on='facility_type')
compas = compas.rename(columns={
'Unnamed 0': 'score',
'pass': '******'
})
compas.pop('zip')
compas.pop('facility_type')
compas['zone'] = compas['zone'].astype(str)
compas['score'] = compas['score'].astype(int)
compas['label_value'] = compas['label_value'].astype(int)
from aequitas.group import Group
from aequitas.bias import Bias
from aequitas.fairness import Fairness
#Group
g = Group()
xtab, attrbs = g.get_crosstabs(compas)
absolute_metrics = g.list_absolute_metrics(xtab)
xtab[[col for col in xtab.columns if col not in absolute_metrics]]
group_df = xtab[['attribute_name', 'attribute_value'] +
[col for col in xtab.columns
if col in absolute_metrics]].round(4)
abs_gpo = xtab[['attribute_name', 'attribute_value'] +
[col for col in xtab.columns
if col in absolute_metrics]].round(4)
#Bias
bias = Bias()
bdf = bias.get_disparity_predefined_groups(xtab,
original_df=compas,
ref_groups_dict={
'zone': 'West',
'facility_group': 'grocery'
},
alpha=0.05)
# View disparity metrics added to dataframe
bias_bdf = bdf[['attribute_name', 'attribute_value'] +
bias.list_disparities(bdf)].round(2)
majority_bdf = bias.get_disparity_major_group(xtab, original_df=compas)
bias_maj_bdf = majority_bdf[['attribute_name', 'attribute_value'] +
bias.list_disparities(majority_bdf)].round(2)
min_bdf = bias.get_disparity_min_metric(xtab, original_df=compas)
bias_min_bdf = min_bdf[['attribute_name', 'attribute_value'] +
bias.list_disparities(min_bdf)].round(2)
min_bdf[['attribute_name', 'attribute_value'] +
bias.list_disparities(min_bdf)].round(2)
#Fairness
fair = Fairness()
fdf = fair.get_group_value_fairness(bdf)
parity_determinations = fair.list_parities(fdf)
fair_fdf = fdf[['attribute_name', 'attribute_value'] + absolute_metrics +
bias.list_disparities(fdf) + parity_determinations].round(2)
gaf = fair.get_group_attribute_fairness(fdf)
fairness_df = fdf.copy()
gof = fair.get_overall_fairness(fdf)
tab_bias_fair = fair_fdf[[
'attribute_name', 'attribute_value', 'for', 'fnr', 'for_disparity',
'fnr_disparity', 'FOR Parity', 'FNR Parity'
]]
tab_bias_fair.rename(columns={
'attribute_value': 'group_name',
'FOR Parity': 'for_parity',
'FNR Parity': 'fnr_parity',
'for': 'for_'
},
inplace=True)
print(tab_bias_fair)
return tab_bias_fair
def audit(df, configs, preprocessed=False):
"""
:param df:
:param configs:
: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,
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,
check_significance=configs.check_significance,
alpha=configs.alpha,
selected_significance=configs.selected_significance,
mask_significance=configs.mask_significance)
elif ref_groups_method == 'majority':
bias_df = b.get_disparity_major_group(
groups_model,
df,
check_significance=configs.check_significance,
alpha=configs.alpha,
selected_significance=configs.selected_significance,
mask_significance=configs.mask_significance)
else:
bias_df = b.get_disparity_min_metric(
df=groups_model,
original_df=df,
check_significance=configs.check_significance,
alpha=configs.alpha,
label_score_ref='fpr',
selected_significance=configs.selected_significance,
mask_significance=configs.mask_significance)
print('Any NaN?: ', bias_df.isnull().values.any())
print('bias_df shape:', bias_df.shape)
aqp = Plot()
if configs.plot_bias_metrics:
if len(configs.plot_bias_metrics) == 1:
fig1 = aqp.plot_group_metric(
bias_df, group_metric=configs.plot_bias_metrics[0])
elif len(configs.plot_bias_metrics) > 1:
fig1 = aqp.plot_group_metric_all(
bias_df, metrics=configs.plot_disparity_attributes)
if (len(configs.plot_bias_disparities) == 1) and (len(
configs.plot_disparity_attributes) == 1):
fig2 = aqp.plot_disparity(
bias_df,
group_metric=configs.plot_bias_disparities[0],
attribute_name=configs.plot_disparity_attributes[0])
elif (len(configs.plot_bias_disparities) > 1) or (len(
configs.plot_disparity_attributes) > 1):
fig2 = aqp.plot_disparity_all(
bias_df,
metrics=configs.plot_bias_disparities,
attributes=configs.plot_disparity_attributes)
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 configs.plot_bias_metrics:
if len(configs.plot_bias_metrics) == 1:
fig3 = aqp.plot_fairness_group(
group_value_df, group_metric=configs.plot_bias_metrics[0])
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) and (len(
configs.plot_disparity_attributes) == 1):
fig4 = aqp.plot_fairness_disparity(
group_value_df,
group_metric=configs.plot_bias_disparities[0],
attribute_name=configs.plot_disparity_attributes[0])
elif (len(configs.plot_bias_disparities) > 1) or (len(
configs.plot_disparity_attributes) > 1):
fig4 = aqp.plot_fairness_disparity_all(
group_value_df,
metrics=configs.plot_bias_disparities,
attributes=configs.plot_disparity_attributes)
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
