Minimax Optimal Regret for Causal Logistic Bandits with Counterfactual Fairness

TL;DR

This paper studies causal logistic bandits with counterfactual fairness constraints and provides matching upper and lower bounds for the expected joint loss. The authors propose an explore--then--exploit procedure and an adaptive algorithm to achieve optimal regret. These algorithms balance exploration and exploitation to minimize regret and cumulative stage-wise positive violations.

Why It Matters

The paper's contribution is significant because it addresses the challenge of ensuring fairness in decision-making systems that learn from feedback. The proposed algorithms can be applied to various real-world scenarios, such as recommendation systems and personalized medicine. By providing a framework for counterfactual fairness, the paper enables the development of more equitable and transparent AI systems.

Key Findings

• identification of a necessary coverage condition
• construction of worst-case families satisfying the full-rank condition
• development of explore--then--exploit and adaptive algorithms with optimal regret guarantees