In [1]:

import sys
import json
from IPython import display
import pandas as pd
import numpy as np
import category_encoders as ce
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split
from sklearn.metrics import ConfusionMatrixDisplay, RocCurveDisplay, precision_score, recall_score
from sklearn.pipeline import Pipeline

# parent directory to work with dev
sys.path.insert(0, '..')
import verifyml.model_card_toolkit as mctlib
from verifyml.model_card_toolkit import model_card_pb2, ModelCard
from verifyml.model_card_toolkit.utils.tally_form import tally_form_to_mc
from verifyml.model_tests.utils import plot_to_str
from verifyml.model_tests.FEAT import (
    SubgroupDisparity,
    MinMaxMetricThreshold,
    Perturbation,
    SHAPFeatureImportance,
    FeatureImportance,
    DataShift
)

import sys import json from IPython import display import pandas as pd import numpy as np import category_encoders as ce from sklearn.ensemble import RandomForestClassifier from sklearn.model_selection import train_test_split from sklearn.metrics import ConfusionMatrixDisplay, RocCurveDisplay, precision_score, recall_score from sklearn.pipeline import Pipeline # parent directory to work with dev sys.path.insert(0, '..') import verifyml.model_card_toolkit as mctlib from verifyml.model_card_toolkit import model_card_pb2, ModelCard from verifyml.model_card_toolkit.utils.tally_form import tally_form_to_mc from verifyml.model_tests.utils import plot_to_str from verifyml.model_tests.FEAT import ( SubgroupDisparity, MinMaxMetricThreshold, Perturbation, SHAPFeatureImportance, FeatureImportance, DataShift )

In [2]:

# Credit card fraud Dataset
df = pd.read_csv("../data/fraud.csv")
x = df.drop("is_fraud", axis=1)
y = df["is_fraud"]


# Train-Test data Split
x_train, x_test, y_train, y_test = train_test_split(
    x, y, test_size=0.5, random_state=50
)


## Build ML model with protected attributes as model features

# Apply one hot encoding to categorical columns (auto-detect object columns) and random forest model in the pipeline
estimator = Pipeline(steps=[('onehot', ce.OneHotEncoder(use_cat_names=True)),
                      ('classifier', RandomForestClassifier(n_estimators=4, max_features="sqrt", random_state = 882))])


# Fit, predict and compute performance metrics
estimator.fit(x_train, y_train)

output = x_test.copy() # x_test df with output columns, to be appended later
y_pred = estimator.predict(x_test)
y_probas = estimator.predict_proba(x_test)[::, 1]

precision_train = round(precision_score(y_train, estimator.predict(x_train)),3)
recall_train = round(recall_score(y_train, estimator.predict(x_train)), 3)
precision_test = round(precision_score(y_test, y_pred),3)
recall_test = round(recall_score(y_test, y_pred), 3)


# Add output columns to this dataframe, to be used as a input for feat tests
output["truth"] = y_test
output["prediction"] = y_pred
output["prediction_probas"] = y_probas


# Dataframe with categorical features encoded
x_train_encoded = estimator[0].transform(x_train)
x_test_encoded = estimator[0].transform(x_test)


# Get feature importance values
df_importance = pd.DataFrame(
    {"features": x_test_encoded.columns, "value": estimator[-1].feature_importances_}
)

# Credit card fraud Dataset df = pd.read_csv("../data/fraud.csv") x = df.drop("is_fraud", axis=1) y = df["is_fraud"] # Train-Test data Split x_train, x_test, y_train, y_test = train_test_split( x, y, test_size=0.5, random_state=50 ) ## Build ML model with protected attributes as model features # Apply one hot encoding to categorical columns (auto-detect object columns) and random forest model in the pipeline estimator = Pipeline(steps=[('onehot', ce.OneHotEncoder(use_cat_names=True)), ('classifier', RandomForestClassifier(n_estimators=4, max_features="sqrt", random_state = 882))]) # Fit, predict and compute performance metrics estimator.fit(x_train, y_train) output = x_test.copy() # x_test df with output columns, to be appended later y_pred = estimator.predict(x_test) y_probas = estimator.predict_proba(x_test)[::, 1] precision_train = round(precision_score(y_train, estimator.predict(x_train)),3) recall_train = round(recall_score(y_train, estimator.predict(x_train)), 3) precision_test = round(precision_score(y_test, y_pred),3) recall_test = round(recall_score(y_test, y_pred), 3) # Add output columns to this dataframe, to be used as a input for feat tests output["truth"] = y_test output["prediction"] = y_pred output["prediction_probas"] = y_probas # Dataframe with categorical features encoded x_train_encoded = estimator[0].transform(x_train) x_test_encoded = estimator[0].transform(x_test) # Get feature importance values df_importance = pd.DataFrame( {"features": x_test_encoded.columns, "value": estimator[-1].feature_importances_} )

Get confusion matrix and ROC curve on train/test set¶

In [3]:

# Train set
ConfusionMatrixDisplay.from_estimator(estimator, x_train, y_train)
confusion_matrix_train = plot_to_str()

RocCurveDisplay.from_estimator(estimator, x_train, y_train)
roc_curve_train = plot_to_str()

# Test set
ConfusionMatrixDisplay.from_estimator(estimator, x_test, y_test)
confusion_matrix_test = plot_to_str()

RocCurveDisplay.from_estimator(estimator, x_test, y_test)
roc_curve_test = plot_to_str()

# Train set ConfusionMatrixDisplay.from_estimator(estimator, x_train, y_train) confusion_matrix_train = plot_to_str() RocCurveDisplay.from_estimator(estimator, x_train, y_train) roc_curve_train = plot_to_str() # Test set ConfusionMatrixDisplay.from_estimator(estimator, x_test, y_test) confusion_matrix_test = plot_to_str() RocCurveDisplay.from_estimator(estimator, x_test, y_test) roc_curve_test = plot_to_str()

Run some FEAT Tests on the data¶

In [4]:

# ROC/Min Max Threshold Test

smt_test = MinMaxMetricThreshold(
    #test_name="",        # Default test name and description will be used accordingly if not specified
    #test_desc="",
    attr="gender",
    metric="fpr",
    threshold=0.025,
    #proba_threshold = 0.6 # Outcome probability threshold, default at 0.5
)
smt_test.run(df_test_with_output=output)
smt_test.plot()

smt_test2 = MinMaxMetricThreshold(
    attr="age",
    metric="fpr",
    threshold=0.025,
)
smt_test2.run(df_test_with_output=output)
smt_test2.plot()

# ROC/Min Max Threshold Test smt_test = MinMaxMetricThreshold( #test_name="", # Default test name and description will be used accordingly if not specified #test_desc="", attr="gender", metric="fpr", threshold=0.025, #proba_threshold = 0.6 # Outcome probability threshold, default at 0.5 ) smt_test.run(df_test_with_output=output) smt_test.plot() smt_test2 = MinMaxMetricThreshold( attr="age", metric="fpr", threshold=0.025, ) smt_test2.run(df_test_with_output=output) smt_test2.plot()

In [5]:

# Subgroup Disparity Test

sgd_test = SubgroupDisparity(
    attr='age',
    metric='fpr',
    method='ratio',
    threshold=1.5,
)
sgd_test.run(output)
sgd_test.plot(alpha=0.05)     # default alpha argument shows 95% C.I bands

sgd_test2 = SubgroupDisparity(
    attr='gender',
    metric='fpr',
    method='ratio',
    threshold=1.5,
)
sgd_test2.run(output)
sgd_test2.plot(alpha=0.05)    # default alpha argument shows 95% C.I bands

# Subgroup Disparity Test sgd_test = SubgroupDisparity( attr='age', metric='fpr', method='ratio', threshold=1.5, ) sgd_test.run(output) sgd_test.plot(alpha=0.05) # default alpha argument shows 95% C.I bands sgd_test2 = SubgroupDisparity( attr='gender', metric='fpr', method='ratio', threshold=1.5, ) sgd_test2.run(output) sgd_test2.plot(alpha=0.05) # default alpha argument shows 95% C.I bands

In [6]:

# Subgroup Perturbation Test

np.random.seed(123)
pmt = Perturbation(
    attr='age',
    metric='fpr',
    method='ratio',
    threshold=1.5,
    #proba_threshold=0.6,  # Outcome probability threshold, default at 0.5
)

pmt.run(
    x_test=x_test,
    y_test=y_test,
    encoder=estimator[0],
    model=estimator[-1]
)

pmt.plot(alpha=0.05)   # default alpha argument shows 95% C.I bands

# Subgroup Perturbation Test np.random.seed(123) pmt = Perturbation( attr='age', metric='fpr', method='ratio', threshold=1.5, #proba_threshold=0.6, # Outcome probability threshold, default at 0.5 ) pmt.run( x_test=x_test, y_test=y_test, encoder=estimator[0], model=estimator[-1] ) pmt.plot(alpha=0.05) # default alpha argument shows 95% C.I bands

In [7]:

pmt.result

pmt.result

Out[7]:

	fpr of original data	fpr of perturbed data	ratio	passed
age_26-39	0.021	0.019	1.140	True
age_40-64	0.024	0.019	1.268	True
age_<=25	0.012	0.024	0.482	True
age_>=65	0.011	0.012	0.959	True

In [8]:

# Shapely Importance Features Test

shap_test = SHAPFeatureImportance(
    attrs=['gender','age'],
    threshold=10
)

shap_test.run(
    model=estimator[-1],
    model_type='trees',
    x_train_encoded=x_train_encoded,
    x_test_encoded=x_test_encoded,
)
shap_test.shap_summary_plot(x_test_encoded)
shap_test.shap_dependence_plot(x_test_encoded, show_all=False) # Show only dependence plots of attributes that failed the test

# Shapely Importance Features Test shap_test = SHAPFeatureImportance( attrs=['gender','age'], threshold=10 ) shap_test.run( model=estimator[-1], model_type='trees', x_train_encoded=x_train_encoded, x_test_encoded=x_test_encoded, ) shap_test.shap_summary_plot(x_test_encoded) shap_test.shap_dependence_plot(x_test_encoded, show_all=False) # Show only dependence plots of attributes that failed the test

model_output = "margin" has been renamed to model_output = "raw"

In [9]:

# User inputted Feature importance test

imp_test = FeatureImportance(
    attrs=['gender','age'],
    threshold=10
)

imp_test.run(df_importance)
imp_test.plot(df_importance, show_n=10)   # Show top 10 most important features

# User inputted Feature importance test imp_test = FeatureImportance( attrs=['gender','age'], threshold=10 ) imp_test.run(df_importance) imp_test.plot(df_importance, show_n=10) # Show top 10 most important features

In [10]:

# Data distribution Shift Test

shift_test = DataShift(
    protected_attr = ['gender','age'],
    method = 'chi2',
    threshold = 0.05
)

shift_test.run(x_train = x_train, x_test = x_test)
shift_test.plot(alpha=0.05)   # default alpha argument shows 95% C.I bands

# Data distribution Shift Test shift_test = DataShift( protected_attr = ['gender','age'], method = 'chi2', threshold = 0.05 ) shift_test.run(x_train = x_train, x_test = x_test) shift_test.plot(alpha=0.05) # default alpha argument shows 95% C.I bands

Bootstrap model card from tally form and scaffold assets¶

We can add the quantitative analysis, explainability analysis and fairness analysis sections to a bootstrap model card for convenience. In this example, we use an existing model card which we created from the tally form response. This is meant only as an example - the dataset and risk evaluation in the model card is a fictional use case.

In [11]:

# Convert form response to model card protobuf
pb = tally_form_to_mc("sample-form-response.json")

# Initialize the mct and scaffold using the existing protobuf
mct = mctlib.ModelCardToolkit(output_dir = "model_card_output", file_name="credit_card_fraud_example")
mc = mct.scaffold_assets(proto=pb)

# Convert form response to model card protobuf pb = tally_form_to_mc("sample-form-response.json") # Initialize the mct and scaffold using the existing protobuf mct = mctlib.ModelCardToolkit(output_dir = "model_card_output", file_name="credit_card_fraud_example") mc = mct.scaffold_assets(proto=pb)

Convert test objects to a model-card-compatible format¶

In [12]:

# init model card test objects
mc_smt_test = mctlib.Test()
mc_smt_test2 = mctlib.Test()
mc_sgd_test = mctlib.Test()
mc_sgd_test2 = mctlib.Test()
mc_pmt_test = mctlib.Test()
mc_shap_test = mctlib.Test()
mc_imp_test = mctlib.Test()
mc_shift_test = mctlib.Test()

# assign tests to them
mc_smt_test.read_model_test(smt_test)
mc_smt_test2.read_model_test(smt_test2)
mc_sgd_test.read_model_test(sgd_test)
mc_sgd_test2.read_model_test(sgd_test2)
mc_pmt_test.read_model_test(pmt)
mc_imp_test.read_model_test(imp_test)
mc_shap_test.read_model_test(shap_test)
mc_shift_test.read_model_test(shift_test)

# init model card test objects mc_smt_test = mctlib.Test() mc_smt_test2 = mctlib.Test() mc_sgd_test = mctlib.Test() mc_sgd_test2 = mctlib.Test() mc_pmt_test = mctlib.Test() mc_shap_test = mctlib.Test() mc_imp_test = mctlib.Test() mc_shift_test = mctlib.Test() # assign tests to them mc_smt_test.read_model_test(smt_test) mc_smt_test2.read_model_test(smt_test2) mc_sgd_test.read_model_test(sgd_test) mc_sgd_test2.read_model_test(sgd_test2) mc_pmt_test.read_model_test(pmt) mc_imp_test.read_model_test(imp_test) mc_shap_test.read_model_test(shap_test) mc_shift_test.read_model_test(shift_test)

In [13]:

# Add quantitative analysis

# Create 4 PerformanceMetric to store our results
mc.quantitative_analysis.performance_metrics = [mctlib.PerformanceMetric() for i in range(0, 4)]
mc.quantitative_analysis.performance_metrics[0].type = "Recall"
mc.quantitative_analysis.performance_metrics[0].value = str(recall_train)
mc.quantitative_analysis.performance_metrics[0].slice = "Training Set"

mc.quantitative_analysis.performance_metrics[1].type = "Precision"
mc.quantitative_analysis.performance_metrics[1].value = str(precision_train)
mc.quantitative_analysis.performance_metrics[1].slice = "Training Set"
mc.quantitative_analysis.performance_metrics[1].graphics.description = (
  'Confusion matrix and ROC Curve')
mc.quantitative_analysis.performance_metrics[1].graphics.collection = [
    mctlib.Graphic(image=confusion_matrix_train), mctlib.Graphic(image=roc_curve_train)
]

mc.quantitative_analysis.performance_metrics[2].type = "Recall"
mc.quantitative_analysis.performance_metrics[2].value = str(recall_test)
mc.quantitative_analysis.performance_metrics[2].slice = "Test Set"

mc.quantitative_analysis.performance_metrics[3].type = "Precision"
mc.quantitative_analysis.performance_metrics[3].value = str(precision_test)
mc.quantitative_analysis.performance_metrics[3].slice = "Test Set"
mc.quantitative_analysis.performance_metrics[3].graphics.description = (
  'Confusion matrix and ROC Curve')
mc.quantitative_analysis.performance_metrics[3].graphics.collection = [
    mctlib.Graphic(image=confusion_matrix_test), mctlib.Graphic(image=roc_curve_test)
]

# Add quantitative analysis # Create 4 PerformanceMetric to store our results mc.quantitative_analysis.performance_metrics = [mctlib.PerformanceMetric() for i in range(0, 4)] mc.quantitative_analysis.performance_metrics[0].type = "Recall" mc.quantitative_analysis.performance_metrics[0].value = str(recall_train) mc.quantitative_analysis.performance_metrics[0].slice = "Training Set" mc.quantitative_analysis.performance_metrics[1].type = "Precision" mc.quantitative_analysis.performance_metrics[1].value = str(precision_train) mc.quantitative_analysis.performance_metrics[1].slice = "Training Set" mc.quantitative_analysis.performance_metrics[1].graphics.description = ( 'Confusion matrix and ROC Curve') mc.quantitative_analysis.performance_metrics[1].graphics.collection = [ mctlib.Graphic(image=confusion_matrix_train), mctlib.Graphic(image=roc_curve_train) ] mc.quantitative_analysis.performance_metrics[2].type = "Recall" mc.quantitative_analysis.performance_metrics[2].value = str(recall_test) mc.quantitative_analysis.performance_metrics[2].slice = "Test Set" mc.quantitative_analysis.performance_metrics[3].type = "Precision" mc.quantitative_analysis.performance_metrics[3].value = str(precision_test) mc.quantitative_analysis.performance_metrics[3].slice = "Test Set" mc.quantitative_analysis.performance_metrics[3].graphics.description = ( 'Confusion matrix and ROC Curve') mc.quantitative_analysis.performance_metrics[3].graphics.collection = [ mctlib.Graphic(image=confusion_matrix_test), mctlib.Graphic(image=roc_curve_test) ]

In [14]:

# You can add the components of a test (e.g. on explainability) in a report
mc.explainability_analysis.explainability_reports = [
    mctlib.ExplainabilityReport(
        type="Top 10 most important features", graphics=mctlib.GraphicsCollection(
            collection = [mctlib.Graphic(name=n, image=i) for n, i in imp_test.plots.items()]
        )
    )
]

# Or you can add it as a test directly
mc.explainability_analysis.explainability_reports.append(
    mctlib.ExplainabilityReport(type="Protected Attributes should not be model's top important features", tests=[mc_shap_test])
)

# You can add the components of a test (e.g. on explainability) in a report mc.explainability_analysis.explainability_reports = [ mctlib.ExplainabilityReport( type="Top 10 most important features", graphics=mctlib.GraphicsCollection( collection = [mctlib.Graphic(name=n, image=i) for n, i in imp_test.plots.items()] ) ) ] # Or you can add it as a test directly mc.explainability_analysis.explainability_reports.append( mctlib.ExplainabilityReport(type="Protected Attributes should not be model's top important features", tests=[mc_shap_test]) )

In [15]:

# The bootstrap template comes with two requirements on fairness analysis:
# Minimum acceptable service and Equal false positive rate
# We add the relevant tests associated with it
mc.fairness_analysis.fairness_reports[0].tests = [mc_smt_test,mc_smt_test2]
mc.fairness_analysis.fairness_reports[1].tests = [mc_sgd_test,mc_sgd_test2]

# We also add a test for attribute shift between the training and testing dataset for additional reliablity check
mc.fairness_analysis.fairness_reports.append(
    mctlib.FairnessReport(type="Distribution of attribute subgroups should be silimiar across different datasets", tests=[mc_shift_test])
)
mc.fairness_analysis.fairness_reports.append(
    mctlib.FairnessReport(type='Fairness metric for subgroups in original data and perturbed data should be similar', tests=[mc_pmt_test])
)

mct.update_model_card(mc)

# The bootstrap template comes with two requirements on fairness analysis: # Minimum acceptable service and Equal false positive rate # We add the relevant tests associated with it mc.fairness_analysis.fairness_reports[0].tests = [mc_smt_test,mc_smt_test2] mc.fairness_analysis.fairness_reports[1].tests = [mc_sgd_test,mc_sgd_test2] # We also add a test for attribute shift between the training and testing dataset for additional reliablity check mc.fairness_analysis.fairness_reports.append( mctlib.FairnessReport(type="Distribution of attribute subgroups should be silimiar across different datasets", tests=[mc_shift_test]) ) mc.fairness_analysis.fairness_reports.append( mctlib.FairnessReport(type='Fairness metric for subgroups in original data and perturbed data should be similar', tests=[mc_pmt_test]) ) mct.update_model_card(mc)

Model Card Display¶

In [16]:

# Export to html
html = mct.export_format(output_file="credit_card_fraud_example.html")
display.display(display.HTML(html))

# Export to html html = mct.export_format(output_file="credit_card_fraud_example.html") display.display(display.HTML(html))

Model Card for Credit Card Fraud Model

Model Details

Overview

Sample example of a risk assessment of a credit card fraud model. Binary prediction problem (fraud or no fraud). Customers flagged as potentially fraudulent will be passed to internal investigation team for follow-up.

Version

name: v1

Owners

Timothy, Product Owner(s)

Swan, Model Developer(s)

Jason, Reviewer(s)

Regulatory requirements

MAS Fairness, Ethics, Accountability and Transparency (FEAT) principles

Considerations

Intended Users

• Credit card fraud team and credit card holders

Use Cases

• Increase accuracy of predicting credit card fraud over the existing rule-based model, saving the bank time and energy for each false positive case and avoiding reputation harm from false negative cases.

Fairness Considerations

• Group at risk: race, age, gender
  Benefits: A more precise model will reduce the number of customers being mistakenly labelled as fraudulent in the existing rules based model, which takes 7 man-days to resolve before a credit card could be unfrozen.
  Harms: Customers who are in the false-positive category will have their credit card frozen and may be excluded from the financial services of the bank for up to 7 days.
  Mitigation Strategy: Because there is less data for certain demographic groups (e.g. youth, elderly), the model can have much higher/lower false-positive rates for that segment than that of others. We will prioritize such cases after the initial model score to add a 2nd level of check and minimise disruption to the customer.

Datasets

Credit Card Dataset

Standard credit card dataset

Sensitive data

• age
• gender

Sensitive data used in model

• age
• gender

Justification

Age and gender is an important predictor of credit card fraud

CRM

Customer information database

Sensitive data

• age
• gender
• race
• religion
• id

Sensitive data used in model

• age
• gender

Justification

Age and gender is an important predictor of credit card fraud

Quantitative Analysis

Recall - 0.936 (Training Set)

Precision - 0.998 (Training Set)

Confusion matrix and ROC Curve

Recall - 0.77 (Test Set)

Precision - 0.955 (Test Set)

Confusion matrix and ROC Curve

Explainability Analysis

Top 10 most important features

Protected Attributes should not be model's top important features

Shapely Feature Importance Test

Description: Test if the subgroups of the protected attributes are the top ranking influential variables under shapely feature importance value. To pass, subgroups should not be ranked in the top 10 features.

Threshold: 10

Result:

	feature_rank	passed
age_>=65	10.0	False
gender_F	13.0	True
gender_M	16.0	True
age_40-64	18.0	True
age_26-39	21.0	True
age_<=25	26.0	True

Failed

Fairness Analysis

Minimum acceptable service

Segment: Age and gender

Description: False positive rate for the credit scoring model should be below 2.5% which is the existing average false positive rate of the rule based model

Min Max Threshold Test

Description: Test if the fpr of the subgroups within gender is lower than the threshold of 0.025.

Threshold: 0.025

Result:

	fpr at current probability threshold	passed
gender_F	0.024	True
gender_M	0.009	True

Passed

Min Max Threshold Test

Description: Test if the fpr of the subgroups within age is lower than the threshold of 0.025.

Threshold: 0.025

Result:

	fpr at current probability threshold	passed
age_26-39	0.021	True
age_40-64	0.024	True
age_<=25	0.012	True
age_>=65	0.011	True

Passed

Equal false positive rate

Segment: Age and gender

Description: Disparity ratio of false positive rates of any 2 bins in the respective attribute should not be more than a factor of 1.5

Subgroup Disparity Test

Description: Test if the maximum ratio of the false postive rate of any 2 groups within age attribute exceeds 1.5. To pass, this value cannot exceed the threshold.

Threshold: 1.5

Result:

	age_fpr_max_ratio
0	2.094

Failed

Subgroup Disparity Test

Description: Test if the maximum ratio of the false postive rate of any 2 groups within gender attribute exceeds 1.5. To pass, this value cannot exceed the threshold.

Threshold: 1.5

Result:

	gender_fpr_max_ratio
0	2.574

Failed

Distribution of attribute subgroups should be silimiar across different datasets

Data Shift Test

Description: Test if there is any shift in the distribution of the attribute subgroups across the different datasets. To pass, the p-value calculated from a chi-square test of independence between the datasets should be greater than 5.0% significance level.

Threshold: 0.05

Result:

	training_distribution	eval_distribution	p-value	passed
gender_F	0.526	0.544	0.234	True
gender_M	0.474	0.456	0.234	True
age_26-39	0.23	0.229	0.798	True
age_40-64	0.323	0.321	0.798	True
age_<=25	0.044	0.05	0.798	True
age_>=65	0.403	0.4	0.798	True

Passed

Fairness metric for subgroups in original data and perturbed data should be similar

Subgroup Perturbation Test

Description: Test if the ratio of the false postive rate of the age subgroups of the original dataset and the perturbed dataset exceeds the threshold. The metric for perturbed dataset will be the denominator. To pass, this computed value cannot exceed 1.5.

Threshold: 1.5

Result:

	fpr of original data	fpr of perturbed data	ratio	passed
age_26-39	0.021	0.019	1.14	True
age_40-64	0.024	0.019	1.268	True
age_<=25	0.012	0.024	0.482	True
age_>=65	0.011	0.012	0.959	True

Passed

In [17]:

# Import model scorecard of second model (without protected attribute as model feature) from external python file
# Same steps taken as the first model, excluding irrelevant tests

# Model card to be compared with that of first model
from model_without_protected_attributes import mc2, mct2

# Import model scorecard of second model (without protected attribute as model feature) from external python file # Same steps taken as the first model, excluding irrelevant tests # Model card to be compared with that of first model from model_without_protected_attributes import mc2, mct2

In [18]:

# Export to html
html2 = mct2.export_format(output_file="credit_card_fraud_example2.html")
display.display(display.HTML(html2))

# Export to html html2 = mct2.export_format(output_file="credit_card_fraud_example2.html") display.display(display.HTML(html2))

Model Card for Credit Card Fraud Model, without protected attributes as model features

Model Details

Overview

Sample example of a risk assessment of a credit card fraud model. Binary prediction problem (fraud or no fraud). Customers flagged as potentially fraudulent will be passed to internal investigation team for follow-up.

Version

name: v1

Owners

Timothy, Product Owner(s)

Swan, Model Developer(s)

Jason, Reviewer(s)

Regulatory requirements

MAS Fairness, Ethics, Accountability and Transparency (FEAT) principles

Considerations

Intended Users

• Credit card fraud team and credit card holders

Use Cases

• Increase accuracy of predicting credit card fraud over the existing rule-based model, saving the bank time and energy for each false positive case and avoiding reputation harm from false negative cases.

Fairness Considerations

• Group at risk: race, age, gender
  Benefits: A more precise model will reduce the number of customers being mistakenly labelled as fraudulent in the existing rules based model, which takes 7 man-days to resolve before a credit card could be unfrozen.
  Harms: Customers who are in the false-positive category will have their credit card frozen and may be excluded from the financial services of the bank for up to 7 days.
  Mitigation Strategy: Because there is less data for certain demographic groups (e.g. youth, elderly), the model can have much higher/lower false-positive rates for that segment than that of others. We will prioritize such cases after the initial model score to add a 2nd level of check and minimise disruption to the customer.

Datasets

Credit Card Dataset

Standard credit card dataset

Sensitive data

• age
• gender

Sensitive data used in model

• age
• gender

Justification

Age and gender is an important predictor of credit card fraud

CRM

Customer information database

Sensitive data

• age
• gender
• race
• religion
• id

Sensitive data used in model

• age
• gender

Justification

Age and gender is an important predictor of credit card fraud

Quantitative Analysis

Recall - 0.958 (Training Set)

Precision - 0.981 (Training Set)

Confusion matrix and ROC Curve

Recall - 0.711 (Test Set)

Precision - 0.858 (Test Set)

Confusion matrix and ROC Curve

Explainability Analysis

Top 10 most important features

Fairness Analysis

Minimum acceptable service

Segment: Age and gender

Description: False positive rate for the credit scoring model should be below 2.5% which is the existing average false positive rate of the rule based model

Min Max Threshold Test

Description: Test if the fpr of the subgroups within gender is lower than the threshold of 0.025.

Threshold: 0.025

Result:

	fpr at current probability threshold	passed
gender_F	0.059	False
gender_M	0.06	False

Failed

Min Max Threshold Test

Description: Test if the fpr of the subgroups within age is lower than the threshold of 0.025.

Threshold: 0.025

Result:

	fpr at current probability threshold	passed
age_26-39	0.061	False
age_40-64	0.059	False
age_<=25	0.052	False
age_>=65	0.06	False

Failed

Equal false positive rate

Segment: Age and gender

Description: Disparity ratio of false positive rates of any 2 bins in the respective attribute should not be more than a factor of 1.5

Subgroup Disparity Test

Description: Test if the maximum ratio of the false postive rate of any 2 groups within age attribute exceeds 1.5. To pass, this value cannot exceed the threshold.

Threshold: 1.5

Result:

	age_fpr_max_ratio
0	1.175

Passed

Subgroup Disparity Test

Description: Test if the maximum ratio of the false postive rate of any 2 groups within gender attribute exceeds 1.5. To pass, this value cannot exceed the threshold.

Threshold: 1.5

Result:

	gender_fpr_max_ratio
0	1.013

Passed

Comparision between 2 model cards¶

In [19]:

# Compare the score cards (on the same content) between model1 and model2
html_compare=mct.compare_model_cards(mc, mc2, export_path='model_card_output/model_cards/credit_card_fraud_comparision.html')
display.display(display.HTML(html_compare))

# Compare the score cards (on the same content) between model1 and model2 html_compare=mct.compare_model_cards(mc, mc2, export_path='model_card_output/model_cards/credit_card_fraud_comparision.html') display.display(display.HTML(html_compare))

Credit Card Fraud Model

Credit Card Fraud Model, without protected attributes as model features

Performance Metrics

Recall - 0.936

Slice: Training Set

Recall - 0.958

Slice: Training Set

Precision - 0.998

Slice: Training Set

Confusion matrix and ROC Curve

Precision - 0.981

Slice: Training Set

Confusion matrix and ROC Curve

Recall - 0.77

Slice: Test Set

Recall - 0.711

Slice: Test Set

Precision - 0.955

Slice: Test Set

Confusion matrix and ROC Curve

Precision - 0.858

Slice: Test Set

Confusion matrix and ROC Curve

Explainability Reports

Top 10 most important features

Top 10 most important features

Fairness Reports

Minimum acceptable service

Segment: Age and gender

Description: False positive rate for the credit scoring model should be below 2.5% which is the existing average false positive rate of the rule based model

Min Max Threshold Test

Description: Test if the fpr of the subgroups within gender is lower than the threshold of 0.025.

Threshold: 0.025

Result:

	fpr at current probability threshold	passed
gender_F	0.024	True
gender_M	0.009	True

Passed

Min Max Threshold Test

Description: Test if the fpr of the subgroups within age is lower than the threshold of 0.025.

Threshold: 0.025

Result:

	fpr at current probability threshold	passed
age_26-39	0.021	True
age_40-64	0.024	True
age_<=25	0.012	True
age_>=65	0.011	True

Passed

Minimum acceptable service

Segment: Age and gender

Description: False positive rate for the credit scoring model should be below 2.5% which is the existing average false positive rate of the rule based model

Min Max Threshold Test

Description: Test if the fpr of the subgroups within gender is lower than the threshold of 0.025.

Threshold: 0.025

Result:

	fpr at current probability threshold	passed
gender_F	0.059	False
gender_M	0.06	False

Failed

Min Max Threshold Test

Description: Test if the fpr of the subgroups within age is lower than the threshold of 0.025.

Threshold: 0.025

Result:

	fpr at current probability threshold	passed
age_26-39	0.061	False
age_40-64	0.059	False
age_<=25	0.052	False
age_>=65	0.06	False

Failed

Equal false positive rate

Segment: Age and gender

Description: Disparity ratio of false positive rates of any 2 bins in the respective attribute should not be more than a factor of 1.5

Subgroup Disparity Test

Description: Test if the maximum ratio of the false postive rate of any 2 groups within age attribute exceeds 1.5. To pass, this value cannot exceed the threshold.

Threshold: 1.5

Result:

	age_fpr_max_ratio
0	2.094

Failed

Subgroup Disparity Test

Description: Test if the maximum ratio of the false postive rate of any 2 groups within gender attribute exceeds 1.5. To pass, this value cannot exceed the threshold.

Threshold: 1.5

Result:

	gender_fpr_max_ratio
0	2.574

Failed

Equal false positive rate

Segment: Age and gender

Description: Disparity ratio of false positive rates of any 2 bins in the respective attribute should not be more than a factor of 1.5

Subgroup Disparity Test

Description: Test if the maximum ratio of the false postive rate of any 2 groups within age attribute exceeds 1.5. To pass, this value cannot exceed the threshold.

Threshold: 1.5

Result:

	age_fpr_max_ratio
0	1.175

Passed

Subgroup Disparity Test

Description: Test if the maximum ratio of the false postive rate of any 2 groups within gender attribute exceeds 1.5. To pass, this value cannot exceed the threshold.

Threshold: 1.5

Result:

	gender_fpr_max_ratio
0	1.013

Passed