Fair Graph Learning
Overview The following two works reduce prediction discrimination by optimizing adjacency matrices, which can improve fairness for link prediction tasks: On Dyadic Fairness: Exploring and Mitigating Bias in Graph Connections （ICLR 2021） All of the Fairness for Edge Prediction with Optimal Transport (AISTATS 2021) 通过修改原图的敏感属性，使用对比学习来实现模型对敏感属性的鲁棒性，即敏感属性的修改不会影响模型的输出: Towards a Unified Framework for Fair and Stable Graph Representation Learning (UAI 2021) 使用对抗训练策略来增强图，使得增强图与敏感属性的关系（MI）最小，基于增强图训练的representation可以实现fairness: Learning Fair Graph Representations via Automated Data Augmentations (ICLR 2023) 证明了message passing的neighbor aggregation会使得拓扑偏差积累到node representation中，在GNN的signal denoising优化框架中加入fairness regularization，使得学习到的节点表示向量要满足，不同敏感group具有相同的期望logits: ...