Nonlinear Graph Learning-Convolutional Networks for Node Classification

摘要

Graph Convolutional Networks have been widely used for node classification. Since the original data usually contains nonlinear relationships that are difficult to capture and includes noise that leads to the poor performance of the constructed graph representation, the paper proposes a novel Nonlinear Graph Learning-Convolutional Network (NGLCN) based on the kernel method and graph representation learning. Specifically, NGLCN first uses a kernel method to map the original data into kernel space, making the original linearly separable to capture the nonlinear relationship between the data, and then uses a feature selection based on structure information to remove the noisy and redundant feature and constructs a high-quality graph representation, and finally employs a common graph convolutional network to conduct node classification tasks. Experimental results on eight benchmark datasets show that NGLCN outperforms the state-of-the-art traditional graph convolutional networks.