Higher-order learning for fairness: a reinforcement learning framework for equitable and efficient influence maximization

Abstract In the governance of online public opinion, the challenge of maximizing influence increasingly requires the simultaneous optimization of both dissemination breadth and fairness among groups. Traditional methods often rely on static network structures, which tend to concentrate influence within dominant groups. To address this issue, we propose the High-Order Learning for Fairness (HOLF) framework. This framework models group fairness as an optimizable graph signal and integrates it comprehensively into the reinforcement learning decision-making process. Specifically: First, we introduce a method for constructing a group-aware hypergraph to encode high-order associations and design a dual-encoder to jointly encode node topological structures and group features. Second, we develop a two-stream Q-network to extract structural influence and group fairness signals separately, employing an adaptive fusion method for balanced evaluation. Finally, we formulate a parameterized reward function, enabling agents to balance dual objectives during sequential decision-making. Experiments conducted on five real-world networks demonstrate that while maintaining an average of 82% of the dissemination efficiency achieved by the greedy algorithm, HOLF reduces group coverage inequality by an average of 16%, consistently outperforming mainstream baseline methods. This work provides an effective solution to the efficiency-fairness trade-off problem in influence maximization and advances fairness-oriented research in information dissemination.