Neighborhood convolutional graph neural network

Jinsong Chen, Boyu Li, Kun He

Abstract: Highlights•Reveal limitations of insufficient modeling capacity and limited scalability of decoupled graph convolutional networks.•Introduce a new paradigm of decoupled GCNs that controls the training cost.•The developed model can adaptively learn node representations of diverse graphs.•Its effectiveness is demonstrated for both homophilic and heterophilic graphs.