Abstract
Graph representation learning has become a vital tool for capturing intricate relationships within networked systems. However, as graphs scale up, data privacy concerns and regulatory restrictions often isolate data within different institutions, hindering collaborative analysis. While federated learning (FL) and swarm learning (SL) offer decentralized solutions, FL relies on a central coordinator, and SL struggles with non-I.I.D. graph data, limiting their flexibility and applicability to complex graph structures. Therefore, establishing a mutual trust mechanism among graph agencies is crucial to unlocking the full potential of graphs. In this article, we introduce a cooperative network learning (CNL) framework to ensure secure graph computing across diverse graph tasks. CNL unifies local and global graph neural network (GNN) perspectives through secure virtual graph connections, enabling collaborative training on decentralized graphs without centralized coordination, while homomorphic encryption and secure communication further preserve privacy. Moreover, CNL enables each agency to maintain autonomy by designing heterogeneous models tailored to its own tasks, going beyond FL and SL. Extensive experiments on contagion dynamics prediction and various traditional graph tasks demonstrate that our CNL architecture outperforms state-of-the-art GNNs developed at individual sites, revealing that CNL provides a reliable, fair, secure, privacy-preserving, and global perspective for effective and personalized network applications. This framework promises to address privacy concerns in graph-related research and integrate decentralized graph data structures to benefit the network research community in cooperation and innovation.