Comparative Study of Gender-Aware Contextual Bandits Performance and Fairness Across Different Gender Groups

Abstract

The study evaluates the performance and fairness of contextual bandit algorithms in gender-aware recommendation systems, utilizing the MovieLens 1M dataset with simulated biased feedback (β = 0.2 for female users). Standard algorithms like Linear Upper Confidence Bound (LinUCB) and Linear Thompson Sampling (LinTS) are compared with fairness-aware variants (Fair-LinUCB, α-Fair, F-UCB) to assess their ability to balance utility, measured by cumulative regret, and fairness, evaluated through reward gaps and exposure inequality across gender groups. The analysis also examines trade-offs between utility and fairness metrics and establishes a reproducible benchmark for gender-fair bandits. Contributions include a standardized evaluation framework, empirical insights into fairness-utility trade-offs through regret curves and Gini coefficients, and practical guidelines for deploying equitable recommendation systems. Results show that fairness-aware algorithms reduce reward gaps by 40–60% compared to baselines, with regret increases ranging from 50% for Fair-LinUCB (regret: 211.77) to over 600% for F-UCB (regret: 997.82). Fair-LinUCB and α-Fair demonstrate optimal trade-offs, balancing equity and efficiency. The study provides a foundation for equitable sequential decision-making, with future directions including intersectional fairness, online validation, and scalable causal methods.