Long-Term Fairness for Real-Time Decision Making: A Constrained Online Optimization Approach.

Area of Science:

• Artificial Intelligence
• Machine Learning Ethics
• Algorithmic Fairness

Background:

• Machine learning (ML) decisions increasingly impact sensitive attributes, necessitating robust fairness and impartiality.
• Real-time decision-making introduces complex fairness needs, including instantaneous and long-term equity.
• Existing methods struggle with time-varying fairness constraints in dynamic systems.

Purpose of the Study:

• To introduce a framework for ensuring long-term fairness in dynamic decision-making systems.
• To address the challenges posed by time-varying fairness constraints.
• To develop an algorithm capable of adapting to evolving fairness requirements.

Main Methods:

• Formulating the decision problem as a constrained online optimization problem.
• Presenting a novel online algorithm, the long-term fairness-aware online learning algorithm (LoTFair).
• Solving the fairness problem dynamically ('on the fly') to accommodate changing constraints.

Main Results:

• LoTFair demonstrates the ability to ensure long-term fairness in real-time decision-making.
• The algorithm achieves overall fairness while maintaining system performance over extended periods.
• It offers a flexible and efficient solution for dynamic fairness constraints.

Conclusions:

• Long-term fairness is achievable in dynamic ML systems with time-varying constraints.
• The LoTFair algorithm provides an effective approach to balancing immediate and future fairness.
• This work advances the development of equitable and reliable AI systems.