Fairness-aware Graph Attention Networks

Graphs can facilitate modeling of various complex systems and the analyses of the underlying relations within them, such as gene networks and power grids. Hence, learning over graphs has attracted increasing attention recently. Specifically, graph neural networks (GNNs) have been demonstrated to achieve state-of-the-art for various learning tasks over graphs. However, it is also shown that the use of graph structures in GNNs results in the amplification of algorithmic bias. Hence, fairness is an essential problem in GNNs. Motivated by this, this study proposes a novel fairness-aware graph attention network (GAT) design. Conventional GAT is one of the most popular and widely utilized GNN structure. Specifically, the present study first carries out a theoretical analysis in order to demonstrate the sources of algorithmic bias in GAT-based learning. Then, a novel fairness-aware GAT design is developed based the theoretical findings. Experimental results on real-world social networks show that the proposed fairness-aware GAT design can improve group fairness measures together with comparable classification accuracy to the conventional GAT for node classification.