Community Detection with Graph Neural Networks

摘要: We study data-driven methods for community detection in graphs. This estimation problem is typically formulated terms of the spectrum certain operators, as well via posterior inference under probabilistic graphical models. Focusing on random graph families such Stochastic Block Model, recent research has unified these two approaches, and identified both statistical computational signal-to-noise thresholds. We embed resulting class algorithms within a generic family neural networks show that they can reach those thresholds purely manner, without access to underlying generative models with no parameter assumptions. The model also tested real datasets, requiring less steps performing significantly better than rigid parametric