A Memetic Algorithm For User Fairness In Recommender Systems

Recommender systems play a significant role in human decisions on e-commerce platforms, streaming services, and social media. However, they often exhibit algorithmic bias that disproportionately disadvantages users with fewer item interactions, who may receive substantially worse recommendations than more active users. Prior work has used Mixed-Integer Linear Programming (MILP) solvers to re-rank recommendations, aiming to minimize the quality gap between user groups. While effective, this approach incurs high computational costs at scale. To address this limitation, we propose a Memetic Algorithm (MA) approach that achieves solutions within 3–5% Normalized Discounted Cumulative Gain (NDCG) and 10–13% F1 relative to MILP quality. Experiments on Amazon review datasets show that the GA produces solutions with competitive recommendation-quality metrics, complies with user-oriented group-fairness constraints, and achieves 18–174× faster execution than Coin-or branch-and-cut (CBC) and 12–36× faster execution than HiGHS.