Fairness dynamics in graph neural networks: a comparative study of graph-structured neural models with and without gradient-based training

Abstract

Graph neural networks (GNNs) are gaining more and more popularity in high stakes domain due to their ability to learn both from features and relationships. Nevertheless, there are concerns regarding how this accuracy centric optimization used by these models will impact fairness when deployed in socially sensitive areas. This work explores the interplay between predictive accuracy and fairness in GNNs when applied in judicial risk assessment system. A comparative study was performed among three canonical architectures such as graph convolutional networks (GCN), graph sample and aggregate (GraphSAGE) and graph attention networks (GAT) under trained and untrained settings on judicial risk assessment dataset. Fairness was evaluated through metrices like demographic parity (DP), equalized opportunity (Eopp), and equalized odds (Eodds) along with predictive performance metrices. Sensitivity analysis was conducted to investigate the effect of graph construction choices and neighborhood sizes in determing fairness and predictive accuracy. Experimental evidences proved that backpropagation improved predictive performance but in tandem fairness degradation happened. Untrained models exhibited lower fairness gap but that is superficial as weak predictive outcome of those models made group differences suppressed. Among the three trained models GAT was able to strike a good balance between accuracy and fairness while increase in neighborhood size caused little bit improvement in fairness via graph smoothing. The novelty of this work lies with its empericial characterization of GNNs under realistic settings. This study emphasizes the fact that how learning methodology, architectural designs, graph formation influence fairness outcomes. This work enlightens how graph-based models can be applied to decision making scenario and encourages embedding of fairness aware training strategies to it.