IC-1441GCN exhibits structural unfairness in transductive node classification, with demographic parity and equal opportunity gaps between nodes connected to and disconnected from the training set
Jiawei Sun, Kailai Li, Ruoxin Chen, Jie LI, Chentao Wu, Yue Ding, Junchi Yan
The paper measures structural fairness by comparing the classification distribution of two node groups: those with direct connections to the training set and those with no path to the training set. On Cora, GCN shows a demographic parity gap of 0.195 and an equal opportunity gap of 0.177. On Wiki-CS, the gaps are larger at 0.478 and 0.432 respectively. The authors attribute this to GCN's inability to propagate training-node information to isolated test nodes, and show that their InterpGNN-GW module reduces these gaps.
Evidence
correlational
Key metric
GCN on Cora: Δsp = 0.195, Δeo = 0.177; GCN on Wiki-CS: Δsp = 0.478, Δeo = 0.432
Caveat
Experiments conducted on small graphs only due to computational impracticability of calculating the node-pair distance matrix on graphs with millions of nodes.