MulStepNET: stronger multi-step graph convolutional networks via multi-power adjacency matrix combination

Abstract Graph convolutional networks (GCNs) have become the de facto approaches and achieved state-of-the-art results for circumventing many real-world problems on graph-structured data. However, these are usually shallow due to over-smoothing of GCNs with layers, which limits expressive power learning graph representations. The current methods solving limitations bottlenecks high complexity parameters. Although Simple Convolution (SGC) reduces parameters, it fails distinguish feature information neighboring nodes at different distances. To tackle limits, we propose MulStepNET, a stronger multi-step network architecture, that can capture more global information, by simultaneously combining neighborhoods information. When compared existing such as GCN MixHop, MulStepNET aggregates distant distances via multi-power adjacency matrix while fitting fewest parameters being computationally efficient. Experiments citation including Pubmed, Cora, Citeseer demonstrate proposed model improves over SGC 2.8, 3.3, 2.1% respectively keeping similar stability, achieves better performance in terms accuracy stability other baselines.