AI Fairness 360: Features, Metrics, and How It Works

How AI Fairness 360 Works

AIF360 usually follows a measure, mitigate, and remeasure process.

• First, the user prepares a dataset and identifies the target outcome. They also define protected attributes when it is lawful and appropriate to use them for fairness testing.
• Next, the user identifies the groups to compare. The toolkit calculates fairness metrics for the original data or model predictions. These results may show whether one group receives favorable outcomes less often or experiences higher error rates.
• The user then selects a mitigation algorithm. The method may transform the training data, change the learning process, or adjust the model’s output. The fairness metrics are calculated again. Teams can then compare fairness, accuracy, and other performance measures.

AIF360 does not decide what “fair” means for an organization. That choice depends on the use case, applicable law, affected people, and the harm being addressed. The documentation also notes that fairness has different definitions for different situations and stakeholders.

Bias Detection Metrics in AI Fairness 360

Different metrics answer different fairness questions.

1. Statistical Parity Difference

This metric compares favorable outcome rates between unprivileged and privileged groups. A result closer to zero generally indicates a smaller gap.

2. Disparate Impact

Disparate impact compares favorable outcome rates as a ratio. A value closer to one generally indicates more similar selection rates between the groups.

3. Equal Opportunity Difference

This metric compares true positive rates. It asks whether qualified members of different groups are equally likely to receive a favorable prediction. A value of zero indicates equal opportunity under this measure.

4. Average Odds Difference

This metric considers differences in both true positive and false positive rates. It can reveal whether a model makes different types of errors across groups. A value of zero indicates equality of odds under the metric.

5. Theil Index

The Theil index measures inequality in model benefits across individuals. It can support fairness analysis at the individual and group levels.

No single metric captures every type of bias. Metrics can also produce different conclusions. Teams should select measures that align with the decision, risk, and fairness objectives.

Bias Mitigation Algorithms in AI Fairness 360

AIF360 groups mitigation techniques into three stages.

Pre-processing Algorithms

These methods modify data before model training.

• Reweighing changes the importance assigned to training examples.
• Disparate Impact Remover edits feature values to improve group fairness while preserving the ordering of values within groups.
• Optimized Preprocessing transforms features and labels under fairness and distortion constraints.
• Learning Fair Representations creates a useful representation while reducing information about protected attributes.

In-processing Algorithms

These methods change the training process.

• Adversarial Debiasing trains a predictor while reducing another model’s ability to infer protected attributes from its predictions.
• Prejudice Remover adds a fairness-related regularisation term to the learning objective.
• Meta Fair Classifier includes fairness constraints within classification.

Post-processing Algorithms

These methods adjust predictions after training.

• Equalized Odds Postprocessing changes outputs to reduce error-rate differences.
• Calibrated Equalized Odds balances calibration, model performance, and fairness conditions.
• Reject Option Classification adjusts uncertain predictions near the decision boundary to improve fairness.

The right method depends on the level of access available. Pre-processing works when teams can change the data. In-processing requires control over training. Post-processing can help when the trained model cannot be changed.

Learn how multi-agent systems, RAG pipelines, and AI-powered workflows are designed and deployed in real organizations with our Applied Agentic AI Course. Through 40+ demos, 10+ guided practices, 7 hands-on projects, and a capstone, you'll gain practical exposure to the technologies shaping the AI-native workplace.