Fairness Assessment and Bias Mitigation with Fairlearn and the Responsible AI Toolbox

“Research on bias, fairness, transparency, and the myriad dimensions of safety now forms a substantial portion of all of the work presented at major AI and machine-learning conferences.”

– Aileen Nielsen, Practical Fairness: Achieving Fair and Secure Data Models

“If and when computer programs attain superhuman intelligence and unprecedented power, should we begin valuing these programs more than we value humans? ... Do humans have some magical spark, in addition to higher intelligence and greater power, which distinguishes them from pigs, chickens, chimpanzees, and computer programs alike? If yes, where did that spark come from, and why are we certain that an AI could never acquire it? If there is no such spark, would there be any reason to continue assigning special value to human life even after computers surpass humans in intelligence and power?”

&...

Fairness metrics

Fairness metrics are critical tools for ensuring that machine learning models are fair and unbiased. These measures allow for the evaluation of classification models and provide insights into whether certain groups are being unfairly favored or discriminated against. Demographic parity and equalized odds are two of the most widely used fairness metrics, both with their own unique approach to measuring fairness. By using these metrics, organizations can better understand how their models perform and take steps to address any biases that may exist.

Demographic parity

Demographic parity is a fairness metric that compares the predictions made between different groups, ignoring the actual true values. This metric is useful in cases where the input data is known to contain biases and the goal is to measure fairness. However, it is important to note that demographic parity only uses the predicted values and discards the information about the true values. It also uses...

Bias and disparity mitigation with Fairlearn

Fairlearn provides several ways to perform bias and disparity mitigation for real-world problems:

• Post-processing methods: This involves adjusting the predictions made by a machine learning model after it has been trained, to reduce bias and disparity. An example of this is the reject option classifier, which allows you to set a threshold for the prediction scores for certain sensitive features. If the threshold is exceeded, the classifier will reject the prediction and instead return a default label.
• Pre-processing methods: This involves transforming the data before training the machine learning model, to reduce bias and disparity. An example of this is CorrelationRemover, which adjusts the non-sensitive features to remove their correlation with the sensitive features, while retaining as much information as possible.
• In-processing methods: This involves modifying the training process of the machine learning model, to reduce...

The Responsible AI Toolbox

The Responsible AI Toolbox9 provides a range of tools and user interfaces to help developers and stakeholders of AI systems to better understand and monitor AI systems. The concept of responsible AI refers to a method of creating, evaluating, and using AI systems in a safe, ethical, and trustworthy way, making informed decisions, and taking responsible actions.

The toolbox includes four visualization widgets to analyze and make decisions about AI models:

• The Responsible AI dashboard brings together various tools from the toolbox to provide a comprehensive view of responsible AI assessment and debugging. With this dashboard, you can identify model errors, understand why they happen, and take steps to address them. Additionally, the causal decision-making capabilities offer valuable insights to stakeholders and customers.
• The Error Analysis dashboard helps identify model errors and identifies groups of data where a model performs poorly.
...

Summary

To summarize, the integration of Fairlearn and the Responsible AI Toolbox provides a comprehensive solution for responsible AI development and deployment, both within Azure as well as open source development. The dashboard brings together the power of several mature Responsible AI tools and libraries, providing a single pane of glass for conducting a holistic responsible assessment, debugging models, and making informed business decisions. With the Error Analysis dashboard, it is possible to identify model errors and discover cohorts of data for which the model underperforms.

The Fairness Assessment dashboard helps identify groups of people that may be disproportionately negatively impacted by an AI system. The Model Interpretability dashboard, powered by InterpretML, explains black-box models and helps users understand their global behavior and the reasons behind individual predictions.

Counterfactual Analysis and Causal Analysis provide actionable insights for data...

References and further reading

1. https://www.reuters.com/article/us-amazon-com-jobs-automation-insight-idUSKCN1MK08G.
2. https://www.health.pa.gov/topics/HealthStatistics/Statistical-Resources/UnderstandingHealthStats/Documents/Confidence_Intervals_for_a_Crude_Rate.pdf.
3. https://sphweb.bumc.bu.edu/otlt/mph-nodules/bs/bs704_confidence_intervals/bs704_confidence_intervals_print.html.
4. https://mlu-explain.github.io/equality-of-odds/.
5. https://fairlearn.org/main/user_guide/assessment/common_fairness_metrics.html#footcite-bickel1975biasinadmissions.
6. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4840791/.
7. The secret bias hidden in mortgage-approval algorithms – Associated Press: https://apnews.com/article/lifestyle-technology-business-race-and-ethnicity-mortgages-2d3d40d5751f933a88c1e17063657586.
8. One-hot encoding is a technique used to transform categorical data into a numerical format. With one-hot encoding, each category is assigned a binary value, where...