Deep Graph Convolutional Neural Network based on Adaptive Sampling and Aggregation Strategy

Graph neural networks learn node embeddings by recursively sampling and aggregating nodes in a graph, while existing methods have a fixed pattern of node sampling and aggregation, and usually only consider direct neighborhood. When sampling high-order neighborhood, it is easy to produce over-smoothing, which affects the model performance and accuracy. Therefore, this paper proposes an adaptive sampling and aggregation deep graph convolutional network (Adaptive Sampling and Aggregation Depth Graph Convolutional Network, ASA-GCN). A sampling strategy based on node importance is designed to adaptively sample high-order neighbors of nodes. Meanwhile, an aggregation strategy based on self-attention mechanism is proposed to adaptively aggregate node information within a given neighborhood. Finally, experimental results on multiple benchmark graph datasets show that the model has superior performance on the node classification task.