Ensuring Minority Group Rights in Social IoT with Fairness-aware Federated Graph Node Classification

Graph neural networks (GNNs) are a powerful tool to predict the categories of IoT nodes for providing diverse high-quality services in the social Internet of Things (SIoT). Intuitively, SIoT could be divided into several communities based on similar interests and mutual trust, whereas interaction between cross-community nodes may intrigue data privacy. With the popularity of federated learning (FL), a global GNN model can be collaboratively trained by exploiting knowledge from multiple parties in a privacy-preserving way. However, since node attributes across different SIoT communities are usually multi-source and heterogeneous, the performance of an optimum global model may exhibit potential bias against minority social groups, i.e, unfairness, thereby undermining the efficiency of node classification and social behavior analysis. To address this problem, we propose PFedNC, a proportional fairness-aware federated learning framework for node classification in SIoT system. Specifically, we incorporate proportional fairness and related Nash bargaining solution into FL optimization objective, which imposes more fairness while maintaining a similar average performance. Moreover, we design a class-balanced re-weighting loss function to mitigate imbalanced class distribution caused by non-IID node distribution. Extensive experiments on two node-level SIoT simulation datasets demonstrate that our PFedNC framework effectively encourages performance fairness among social groups and achieves a favorable tradeoff between fairness and efficiency.