Core Concepts
This paper proposes a novel framework named FairGI that simultaneously achieves group fairness and individual fairness within groups in graph learning, while maintaining comparable prediction performance.
Abstract
The paper addresses the problem of achieving both group fairness and individual fairness within groups in graph neural networks (GNNs). Existing work on fair graph learning has focused on either group fairness or individual fairness, but not both simultaneously.
The key contributions are:
Introduction of a novel problem concerning the achievement of both group fairness and individual fairness within groups in graph learning.
Proposal of a new metric to measure individual fairness within groups for graphs.
Development of an innovative framework FairGI to ensure group fairness and individual fairness within groups in graph learning while maintaining comparable model prediction performance.
Comprehensive experiments on real-world datasets demonstrating the effectiveness of FairGI in eliminating both group and individual fairness biases while maintaining comparable prediction performance.
The framework consists of three main components:
Individual fairness within groups module: Proposes a novel loss function to minimize bias among individuals within the same group.
Group fairness module: Incorporates adversarial learning and covariance constraint loss functions to optimize for both Equal Opportunity and Statistical Parity.
GNN classifier for node prediction.
The experimental results show that FairGI outperforms state-of-the-art methods in terms of group fairness and individual fairness within groups, while maintaining comparable prediction accuracy. Interestingly, even though FairGI only constrains individual fairness within groups, it achieves the best population individual fairness compared to the baselines.
Stats
The prediction accuracy of our method is comparable to or better than the baselines.
Our method achieves the lowest maximum individual unfairness (MaxIG) across all groups compared to the baselines.
Our method achieves the lowest Statistical Parity (SP) and Equal Opportunity (EO) compared to the baselines.