Abstract
Graph Neural Networks (GNNs) have been widely used in recommender systems due to their ability to model high-order user-item interactions. However, their message-passing mechanism can inadvertently amplify data biases, which may increase group unfairness. Existing methods often focus on suppressing sensitive attributes at the node level, but such constraints alone cannot fully remove biases arising from the graph topology. Both node attribute and topology biases can propagate and amplify through GNN message passing, making it difficult to ensure fair recommendations. To address this, we propose Fair Dual-path Alignment (FairDA), a novel fairness optimization method. FairDA aligns user embeddings learned by LightGCN on the original data (student representations) with fair embeddings obtained from data with sensitive attributes removed. This alignment reduces biases in both node attributes and graph topology, producing low-bias user representations suitable for fair recommendation. We further introduce an information bottleneck-based mutual information constraint. This constraint preserves collaborative filtering signals while removing sensitive information, enhancing fairness without compromising recommendation accuracy. Finally, FairDA dynamically adjusts the loss weights of similar item pairs to regulate inter-group item relationships, further reducing group bias. Extensive experiments on two real-world datasets show that FairDA achieves a superior balance between recommendation accuracy and fairness.