In decentralized federated learning, clients train on their own data and can develop complementary expertise, but each query is typically evaluated only by the client where it arrives. The authors of a new arXiv preprint propose treating inference as a collaborative routing problem, allowing a query to be referred from client to client along the existing network graph.
Their framework, EdgeDRM, gives each client a learned density-based routing potential. When a query arrives, the difference between a client's potential and its neighbors' potentials determines where to refer next. The route is constrained by a budget, and only the clients visited along the way contribute to the final prediction. Both training and inference rely only on neighbor-local information, so no central router or graph-wide client scoring is needed.
The paper establishes estimation guarantees for the learned comparisons and characterizes how local improvement and graph structure govern prior-weighted density gains from referral. Experiments on benchmark and real-world medical data show that EdgeDRM outperforms graph-matched decentralized baselines in most settings, with the clearest benefits under sparse connectivity.