Abstract:Graph Convolutional Network (GCN) has been widely used in graph learningtasks. However, GCN-based models (GCNs) is an inherently coupled trainingframework repetitively conducting the complex neighboring aggregation, whichleads to the limitation of flexibility in processing large-scale graph. Withthe depth of layers increases, the computational and memory cost of GCNs growexplosively due to the recursive neighborhood expansion. To tackle theseissues, we present Node2Grids, a flexible uncoupled training framework thatleverages the independent mapped data for obtaining the embedding. Instead ofdirectly processing the coupled nodes as GCNs, Node2Grids supports a moreefficacious method in practice, mapping the coupled graph data into theindependent grid-like data which can be fed into the efficient ConvolutionalNeural Network (CNN). This simple but valid strategy significantly saves memoryand computational resource while achieving comparable results with the leadingGCN-based models. Specifically, by ranking each node's influence throughdegree, Node2Grids selects the most influential first-order as well assecond-order neighbors with central node fusion information to construct thegrid-like data. For further improving the efficiency of downstream tasks, asimple CNN-based neural network is employed to capture the significantinformation from the mapped grid-like data. Moreover, the grid-level attentionmechanism is implemented, which enables implicitly specifying the differentweights for neighboring nodes with different influences. In addition to thetypical transductive and inductive learning tasks, we also verify our frameworkon million-scale graphs to demonstrate the superiority of the proposedNode2Grids model against the state-of-the-art GCN-based approaches.

Non-recursive graph convolutional networks

Neighborhood Convolutional Graph Neural Network

Neighbor enhanced graph convolutional networks for node classification and recommendation

Neighborhood Convolutional Network: A New Paradigm of Graph Neural Networks for Node Classification

Adaptive Sampling Towards Fast Graph Representation Learning

Going Deep: Graph Convolutional Ladder-Shape Networks

Hierarchical Graph Convolutional Networks for Semi-supervised Node Classification

A deep graph convolutional neural network architecture for graph classification

Efficient Nodes Representation Learning with Residual Feature Propagation

Graph Neural Networks with Non-Recursive Message Passing

Simplifying Graph Convolutional Networks

A simple and effective convolutional operator for node classification without features by graph convolutional networks

An Uncoupled Training Architecture for Large Graph Learning

Locality preserving dense graph convolutional networks with graph context-aware node representations

Robust graph learning with graph convolutional network

Variational Gridded Graph Convolution Network for Node Classification

A Subgraph Sampling Method for Training Large-Scale Graph Convolutional Network.

Multi-scale Graph Convolutional Networks with Self-Attention

Graph Convolutional Networks with Adaptive Neighborhood Awareness

MobileGCN applied to low-dimensional node feature learning