Use AutoAI to train fair models.ipynb

Kernel: Python 3 (ipykernel)

Bias detection and mitigation in AutoAI

This notebook contains the steps and code to demonstrate support of AutoAI experiments with bias detection/mitigation in Watson Machine Learning service. It introduces commands for data retrieval, training experiments, persisting pipelines, testing pipelines and scoring.

Some familiarity with Python is helpful. This notebook uses Python 3.11.

Learning goals

The learning goals of this notebook are:

• Work with Watson Machine Learning experiment to train AutoAI models with bias detection and mitigation.

• Compare trained models quality and fairness.

Contents

This notebook contains the following parts:

1. Setup

2. Optimizer definition

3. Bias detection and mitigation

4. Inspection of pipelines

5. Cleanup

6. Summary and next steps

1. Set up the environment

Before you use the sample code in this notebook, you must perform the following setup tasks:

• Contact with your Cloud Pack for Data administrator and ask him for your account credentials

Install and import the ibm-watsonx-ai and dependecies

Note: ibm-watsonx-ai documentation can be found here.

!pip install wget | tail -n 1
!pip install -U autoai-libs | tail -n 1
!pip install -U ibm-watsonx-ai | tail -n 1
!pip install scikit-learn==1.3.0 | tail -n 1
!pip install -U 'lale[fairness]>=0.8.2' | tail -n 1

Connection to WML

Authenticate the Watson Machine Learning service on IBM Cloud Pack for Data. You need to provide platform url, your username and api_key.

username = 'PASTE YOUR USERNAME HERE'
api_key = 'PASTE YOUR API_KEY HERE'
url = 'PASTE THE PLATFORM URL HERE'

from ibm_watsonx_ai import Credentials

credentials = Credentials(
    username=username,
    api_key=api_key,
    url=url,
    instance_id="openshift",
    version="5.0"
)

Alternatively you can use username and password to authenticate WML services.

credentials = Credentials(
    username=***,
    password=***,
    url=***,
    instance_id="openshift",
    version="5.0"
)

Working with spaces

First of all, you need to create a space that will be used for your work. If you do not have space already created, you can use {PLATFORM_URL}/ml-runtime/spaces?context=icp4data to create one.

• Click New Deployment Space

• Create an empty space

• Go to space Settings tab

• Copy space_id and paste it below

Tip: You can also use SDK to prepare the space for your work. More information can be found here.

Action: Assign space ID below

from ibm_watsonx_ai import APIClient

client = APIClient(credentials)

space_id = 'PASTE YOUR SPACE ID HERE'

You can use list method to print all existing spaces.

client.spaces.list(limit=10)

client.set.default_space(space_id)

'SUCCESS'

Optimizer definition

Training data sets

Define connection information to training data CSV file.

Download training data from git repository.

import wget
import os

filename = 'german_credit_data_biased_training.csv'

url = "https://github.com/IBM/watson-machine-learning-samples/raw/master/cpd5.0/data/bias/german_credit_data_biased_training.csv"
if not os.path.isfile(filename): 
    wget.download(url)

Create data asset

asset_details = client.data_assets.create('german_credit_data_biased_training.csv', filename)
asset_details

Creating data asset...
SUCCESS

{'metadata': {'space_id': 'e8402072-aee5-4fee-8e4f-836a1453ba6d',
  'usage': {'last_updated_at': '2024-05-21T11:25:11Z',
   'last_updater_id': '1000331001',
   'last_update_time': 1716290711056,
   'last_accessed_at': '2024-05-21T11:25:11Z',
   'last_access_time': 1716290711056,
   'last_accessor_id': '1000331001',
   'access_count': 0},
  'rov': {'mode': 0,
   'collaborator_ids': {},
   'member_roles': {'1000331001': {'user_iam_id': '1000331001',
     'roles': ['OWNER']}}},
  'name': 'german_credit_data_biased_training.csv',
  'description': '',
  'asset_type': 'data_asset',
  'origin_country': 'us',
  'resource_key': 'german_credit_data_biased_training.csv',
  'rating': 0.0,
  'total_ratings': 0,
  'catalog_id': '82c126ea-17a0-4742-a6cd-08c712cd5473',
  'created': 1716290711056,
  'created_at': '2024-05-21T11:25:11Z',
  'owner_id': '1000331001',
  'size': 0,
  'version': 2.0,
  'asset_state': 'available',
  'asset_attributes': ['data_asset'],
  'asset_id': 'e1c15f7e-c7c5-44b5-b5ef-89019f94e3f2',
  'asset_category': 'USER',
  'creator_id': '1000331001',
  'guid': 'e1c15f7e-c7c5-44b5-b5ef-89019f94e3f2',
  'href': '/v2/assets/e1c15f7e-c7c5-44b5-b5ef-89019f94e3f2?space_id=e8402072-aee5-4fee-8e4f-836a1453ba6d',
  'last_updated_at': '2024-05-21T11:25:11Z'},
 'entity': {'data_asset': {'mime_type': 'text/csv'}}}

client.data_assets.get_id(asset_details)

'e1c15f7e-c7c5-44b5-b5ef-89019f94e3f2'

from ibm_watsonx_ai.helpers import DataConnection

german_credit_risk = DataConnection(data_asset_id=client.data_assets.get_id(asset_details))

training_data_reference=[german_credit_risk]

Bias detection and mitigation

Terms and definitions:

Fairness Attribute - Bias or fairness is typically measured using some fairness attribute such as Gender, Ethnicity, Age, etc.

Monitored/Reference Group - Monitored group are those values of fairness attribute for which we want to measure bias. The rest of the values of the fairness attributes are called as reference group. In case of Fairness Attribute=Gender, if we are trying to measure bias against females, then Monitored group is “Female” and Reference group is “Male”.

Favourable/Unfavourable outcome - An important concept in bias detection is that of favourable and unfavourable outcome of the model. E.g., Claim approved can be considered as a favourable outcome and Claim denied can be considered as an unfavourable outcome.

Disparate Impact - metric used to measure bias (computed as the ratio of percentage of favourable outcome for the monitored group to the percentage of favourable outcome for the reference group). Bias is said to exist if the disparate impact value is below some threshold.

Optimizer configuration

Provide input information for AutoAI optimizer:

• name - experiment name

• prediction_type - type of the problem

• prediction_column - target column name

• fairness_info - bias detection configuration

• scoring - accuracy_and_disparate_impact combined optimization metric for both accuracy and fairness. For regression learning problem the r2_and_disparate_impact metric is supported (combines r2 and fairness).

fairness_info definition:

• protected_attributes (list of dicts) – subset of features for which fairness calculation is desired.

  • feature - name of feature for which reference_group and monitored_group are specified.

  • reference_group and monitored_group - monitored group are those values of fairness attribute for which we want to measure bias. The rest of the values of the fairness attribute are reference group.

• favorable_labels and unfavorable_labels – label values which are considered favorable (i.e. “positive”). unfavorable_labels are required when prediction type is regression.

Examples of supported configuration:

fairness_info = {
            "protected_attributes": [
                {"feature": "Age", "reference_group": [[26, 26], [30, 75]], 
                                    "monitored_group": [[18, 25], [27, 29]]}
            ],
            "favorable_labels": ["No Risk"]
            }

fairness_info = {
            "protected_attributes": [
                {"feature": "sex", "reference_group": ['male', 'not specified'], 
                                   "monitored_group": ['female']},
                {"feature": "age", "reference_group": [[26, 100]], "monitored_group": [[18, 25], [27, 29]]}
            ],
            "favorable_labels": [[5000.01, 9000]],
            "unfavorable_labels": [[0, 5000], [9000, 1000000]]
            }

fairness_info = {
            "protected_attributes": [
                {"feature": "Sex", "reference_group": ['male'], "monitored_group": ['female']},
                {"feature": "Age", "reference_group": [[26, 75]], "monitored_group": [[18, 25]]}
            ],
            "favorable_labels": ["No Risk"],
            "unfavorable_labels": ["Risk"],
}

from ibm_watsonx_ai.experiment import AutoAI

experiment = AutoAI(credentials, space_id=space_id)

pipeline_optimizer = experiment.optimizer(
    name='Credit Risk Prediction and bias detection - AutoAI',
    prediction_type=AutoAI.PredictionType.BINARY,
    prediction_column='Risk',
    scoring='accuracy_and_disparate_impact',
    fairness_info=fairness_info,
    max_number_of_estimators = 1,
    retrain_on_holdout=False,
    include_only_estimators=["XGBClassifier"]
   )

Experiment run

Call the fit() method to trigger the AutoAI experiment. You can either use interactive mode (synchronous job) or background mode (asychronous job) by specifying background_model=True.

run_details = pipeline_optimizer.fit(
            training_data_reference=training_data_reference,
            background_mode=False)

Training job aa651e79-447e-4199-bbed-588951d2f2a4 completed: 100%|████████| [02:04<00:00,  1.25s/it]

You can use the get_run_status() method to monitor AutoAI jobs in background mode.

Get selected pipeline model

Download and reconstruct a scikit-learn pipeline model object from the AutoAI training job.

experiment_summary = pipeline_optimizer.summary()
experiment_summary.head()

Visualize pipeline

pipeline_name = experiment_summary.index[experiment_summary.holdout_disparate_impact.argmax()]
best_pipeline = pipeline_optimizer.get_pipeline(pipeline_name=pipeline_name)
best_pipeline.visualize()

Each node in the visualization is a machine-learning operator (transformer or estimator). Each edge indicates data flow (transformed output from one operator becomes input to the next). The input to the root nodes is the initial dataset and the output from the sink node is the final prediction. When you hover the mouse pointer over a node, a tooltip shows you the configuration arguments of the corresponding operator (tuned hyperparameters). When you click on the hyperlink of a node, it brings you to a documentation page for the operator.

Test pipeline model locally

Read the data

X_train, X_holdout, y_train, y_holdout = pipeline_optimizer.get_data_connections()[0].read(with_holdout_split=True)

Calculate metrics

For detail description of used metrics you can check the documentation:

• accuracy

• disparate_impact

• accuracy and disparate impact

from lale.lib.aif360 import disparate_impact, accuracy_and_disparate_impact
from sklearn.metrics import accuracy_score

predicted_y = best_pipeline.predict(X_holdout.values)
disparate_impact_scorer = disparate_impact(**fairness_info)
accuracy_disparate_impact_scorer = accuracy_and_disparate_impact(**fairness_info)

print("Accuracy: {:.2f}".format(accuracy_score(y_true= y_holdout, y_pred=predicted_y)))
print("Disparate impact: {:.2f}".format(disparate_impact_scorer(best_pipeline, X_holdout, y_holdout)))
print("Accuracy and disparate impact: {:.2f}".format(accuracy_disparate_impact_scorer(best_pipeline, X_holdout, y_holdout)))

Accuracy: 0.81
Disparate impact: 1.46
Accuracy and disparate impact: 0.75

---

Fairness insights

You can analize favorable outcome distributions using visualize method from utils module.

from ibm_watsonx_ai.utils.autoai.fairness import visualize

visualize(run_details, pipeline_name)

MIME type unknown not supported

MIME type unknown not supported

---

Clean up

If you want to clean up all created assets:

• experiments

• trainings

• pipelines

• model definitions

• models

• functions

• deployments

please follow up this sample notebook.

Summary and next steps

You successfully completed this notebook!

As a next step you can deploy and score the model: Sample notebook.

Check out our Online Documentation for more samples, tutorials, documentation, how-tos, and blog posts.

Authors

Lukasz Cmielowski, PhD, is an Automation Architect and Data Scientist at IBM with a track record of developing enterprise-level applications that substantially increases clients' ability to turn data into actionable knowledge.

Dorota Lączak, software engineer in Watson Machine Learning at IBM

Szymon Kucharczyk, software engineer in Watson Machine Learning at IBM

Copyright © 2021-2025 IBM. This notebook and its source code are released under the terms of the MIT License.