A Federated Learning Approach for Graph Convolutional Neural Networks

In this paper we cast the problem of training a graph neural network based on labeled graph data in a “federated learning” scenario where different agents have access to data from a subset of the network nodes. The learning problem is not decomposable, therefore it does not lend itself to a straightforward mapping onto a distributed multi-agent protocol. We propose a multi-agent federated learning scheme which leverages the local and sparse structure of graph filters to limit the information sharing while emulating the performance of centralized training. Even though we preserve data locality and agent communication is restricted to the neighborhood level, the proposed method still converges in simulation.