Recently, graph neural networks have achieved good performance in graph representation learning. However, most graph neural networks only utilize node low-frequency signals and destroy node similarity when aggregating graph structure and node features, which limits their ability to represent graph-structured data. Therefore, we propose a node similarity preserving graph convolutional network based on full-frequency information (FSP-GCN). It extracts relevant information to the greatest extent from graph structure and node features while preserving node similarity for aggregation. Precisely, to...
