Visualizing and understanding graph convolutional network

• DOI: 10.1007/s11042-020-09885-4
• Publication Date: 2020-11-03T01:03:06Z
• Authors: Xiaoying Shi; Fanshun Lv; Dewen Seng; Jiaming Zhang; Jing Chen; Baixi Xing
• Keywords: 0202 electrical engineering, electronic engineering, information engineering; 02 engineering and technology

The graph convolutional network (GCN), which can handle graph-structured data, is enjoying great interest in recent years. However, while GCN achieved remarkable results for different kinds of tasks, the source of its performance and the underlying decision process remain poorly understood. In this paper, we propose a visual analytics system that supports progressive analysis of GCN executing process and the effect of graph convolution operation. Multiple coordinated views are designed to show the influence of hidden layer parameters, the change of loss/accuracy and activation distributions, and the diffusion process of correctly predicted nodes. In particular, since the traditional t-SNE and force-directed layout methods are unable to show the graph-structured data well, we propose to utilize ‘graphTSNE’, a novel visualization technique for graph-structured data, to present the node layout in a clearer way. The real-world graph dataset is used to demonstrate the usability and effectiveness of our system through case studies. The results manifest that our system can provide sufficient guidance for understanding the working principle of graph convolutional network.