Auditing Fairness Online through Interactive Refinement

The Need

In the era of machine learning, high-stakes decisions are increasingly being made by black box models, leading to concerns about accountability and fairness. These models can exhibit inherent biases, raising the need for a system that ensures accountability and fairness in decision-making processes. While previous efforts have focused on fairness monitoring, specifying appropriate fairness metrics remains challenging. There's a clear need for a solution that not only improves the bounds for fairness metrics but also provides an interactive and iterative process for defining and monitoring fairness specifications, all while meeting governance and regulatory requirements.

The Technology

Our innovative technology, known as AVOIR, addresses the need for accountable and fair decision-making in black box machine learning models by offering the following key features:

  • Automated Inference-Based Optimization: AVOIR utilizes automated inference-based optimization to improve fairness metrics and provides probabilistic guarantees for fairness grammars, enhancing the confidence with which specification violations are reported.
  • Interactive Visual Analysis: A novel visualization mechanism within AVOIR allows users to investigate the context of reported fairness violations, aiding in the refinement of fairness specifications.

Commercial Applications

  • Regulatory Compliance: AVOIR can be used to ensure that machine learning models comply with regulatory requirements related to fairness and accountability, particularly in industries like finance and healthcare.
  • Ethical AI Development: Organizations developing AI solutions can integrate AVOIR to assess and improve the fairness of their models, enhancing ethical AI practices.
  • Risk Mitigation: AVOIR helps mitigate the risks associated with biased decision-making, reducing the potential for legal and reputational damage.


  • Enhanced Fairness: AVOIR improves fairness metrics, ensuring that machine learning models make decisions that are more equitable and less prone to bias.
  • Probabilistic Guarantees: The technology provides probabilistic guarantees for fairness grammars, increasing confidence in the accuracy of fairness violation reports.
  • Interactive Refinement: AVOIR's interactive visual analysis facilitates the refinement of fairness specifications, allowing users to make informed decisions about model fairness.
  • Regulatory Compliance: By using AVOIR, organizations can demonstrate their commitment to fairness and accountability, meeting regulatory requirements and avoiding potential legal consequences.

Loading icon