A transferability benchmark of graph neural networks for cross-domain risk contagion identification: GAT, GraphSAGE, and GCN across financial, cyber, and epidemiological networks

Abstract

Risk contagion across heterogeneous networks—shock propagation between financial markets, lateral movement of intrusions across enterprise hosts, and pathogen spreading along contact graphs—shares a common formal structure but is rarely studied jointly. We construct a unified benchmark that aligns three publicly derived graph corpora (an interbank-style transaction subgraph, an authentication-flow subgraph from enterprise telemetry, and an SNAP contact graph augmented with synthetic exposure labels) under a single node-classification protocol, and run a controlled comparison of three widely deployed graph neural networks: Graph Convolutional Networks (GCN), GraphSAGE, and Graph Attention Networks (GAT). Models are evaluated in three regimes: in-domain training, cold-start cross-domain transfer with frozen encoders, and adapter-based fine-tuning on a 5% target sample. Across 27 (model × source × target) configurations, GAT attains the highest in-domain F1 and the most stable transfer in financial→cyber pairs, GraphSAGE leads on the epidemiological target due to neighborhood-sampling robustness on long-tailed degree distributions, and GCN shows the steepest performance gap when transferred without adaptation. We discuss when neighborhood attention helps versus when sampling-based aggregation suffices, and present an analysis of feature drift and label-noise sensitivity. The benchmark, while not exhaustive, offers a reproducible reference point for cross-domain GNN deployment in risk-sensitive applications.