SpreadGNN: Decentralized Multi-Task Federated Learning for Graph Neural Networks on Molecular Data

Graph Neural Networks (GNNs) are the first choice methods for graph machine learning problems thanks to their ability learn state-of-the-art level representations from graph-structured data. However, centralizing a massive amount of real-world data GNN training is prohibitive due user-side privacy concerns, regulation restrictions, and commercial competition. Federated Learning de-facto standard collaborative models over many distributed edge devices without need centralization. Nevertheless, neural networks in federated setting vaguely defined brings statistical systems challenges. This work proposes SpreadGNN, novel multi-task framework capable operating presence partial labels absence central server time literature. We provide convergence guarantees empirically demonstrate efficacy our on variety non-I.I.D. graph-level molecular property prediction datasets with labels. Our results show that SpreadGNN outperforms trained server-dependent system, even constrained topologies.