Abstract:The research on node classification is based on node embeddings. Node classification accuracy can be improved if the embeddings of different nodes are well discriminated. With the rapid development of deep learning, researchers have proposed many graph neural network models (GNNs), such as GCN and GAT, which generally obtain node embeddings by aggregating neighborhood information. However, such methods only emphasize feature aggregation in neighborhoods and do not consider the class labels of nodes, which leads to the oversmoothing problem and weak differences in inter-class nodes. In this paper, we propose a gated graph attention network based on dual graph convolution for node embedding (GGAN-DGC). To strengthen the embedding difference of inter-class nodes, GGAN-DGC introduces a gated attention mechanism. This mechanism utilizes a supervised gated attention (GA) matrix to separate the GNN aggregation process according to the node class, so as to heterogenize the homogenous graphs. The GA matrix is obtained by the dual graph convolutional network (DGC), which can improve the receptive field of the original graph. In addition, GGAN-DGC adopts triplet loss as the global supervision function of node embedding, which can streng-then the class correlation of node embedding at the global level. Finally, based on the obtained node embedding, nodes can be classified correctly. The experimental results on five datasets confirm that our GGAN-DGC model performs better than other representative methods in node classification, especially for datasets with strong heterophily. In addition, we verify that GGAN-DGC can also perform better than other methods in graph classification experiments.

DualGCN: Representation Learning for Property Graphs Via Dual-level Convolutional Networks

Graph Learning in Low Dimensional Space for Graph Convolutional Networks

Dual Graph Complementary Network

DAGCN: Dual Attention Graph Convolutional Networks

Dual Feature Interaction-Based Graph Convolutional Network

A gated graph attention network based on dual graph convolution for node embedding

Attention-Based Relational Graph Convolutional Network for Knowledge Graph Reasoning

2 Related Work Following the success of deep neural networks for image classification

DPGNN: Dual-perception graph neural network for representation learning

A Novel Representation Learning for Dynamic Graphs Based on Graph Convolutional Networks

Gated Relational Graph Neural Network For Semi-Supervised Learning On Knowledge Graphs

DynGCN: A Dynamic Graph Convolutional Network Based on Spatial-Temporal Modeling

HDGCN: Dual-Channel Graph Convolutional Network with Higher-Order Information for Robust Feature Learning

Dual Graph Representation Learning

Dual Graph Convolutional Network for Semantic Segmentation.

Dual-channel deep graph convolutional neural networks

Neighborhood Convolutional Graph Neural Network

DyGCN: Dynamic Graph Embedding with Graph Convolutional Network

Multi-scale Graph Convolutional Networks with Self-Attention

Graph Convolutional Network For Semi-supervised Node Classification With Subgraph Sketching

Multi-Relational Graph Representation Learning with Bayesian Gaussian Process Network.