A Survey on Graph Convolutional Neural Network

In the past few years, convolutional neural network has attracted widespread attention due to its powerful modeling capabilities, and has achieved great improvement in areas such as natural language processing and computer vision. Traditional convolutional neural network can only process Euclidean data. However, many real-life scenarios, such as transportation networks, social networks and citation networks, are located in the form of graph data. The method used to process graph data previously is network embedding. Specifically, the researchers model the graph-based mission into a two-stage model. In the first stage, a fixed-length representation is learned for each node via capturing the proximity over nodes, this representation is fed into the second stage to solve downstream tasks, e.g., link prediction, node classification and graph classification. In recent years, the powerful modeling capabilities of convolutional neural network and the ubiquity of graph data have inspired researchers to transfer convolutional neural network to graphs, which can solve the graph-based task via an end-to-end manner. The core of graph convolutional neural network is the construction of graph convolution operator and pooling operator. In this paper, we review the graph convolutional neural network. Firstly, the background of graph convolutional neural network and its classical methods are introduced, including spectral methods and spatial methods. The lack of translation invariance on the graph data makes it difficult to define the graph convolution operator. The spectral methods define the convolution in the spectral domain via the convolution theorem, while the spatial methods implement the graph convolution by defining the node correlation in the node domain. Then, the latest developments are introduced. Recent researchers focus on how to model the complicated information on graph via graph convolution neural network, e.g., heterogeneous connection and high-order connection. In addition, how to construct graph convolution neural network on large-scale network also attracts much attention. Moreover, we conclude the graph convolution neural network in many applications, including traffic prediction and recommender system. Finally, the developing trend of graph convolutional neural network is summarized and forecasted. © 2020, Science Press. All right reserved.