Venusaur –One Stitched Toy Graph Neural Network

Shanzai CL

Published on: 2021-09-12

Abstract

Graph Convolutional Network (GCNS) is a powerful deep learning method for graph structure data. Recently, GCNS and its subsequent variants have shown excellent performance in various application fields on real data sets. Despite the success, most current GCN models are very shallow due to the problem of over smoothing. Moreover, there are limitations of planarization, so the graph cannot be characterized in layers. In real applications, a lot of graph information is represented by hierarchical levels, such as maps, conceptual graphs, and flowcharts. Capturing hierarchical information can represent graphs more completely and efficiently. This paper studies the design and analysis of deep image convolutional networks. We proposed the VENUSAUR model, using three simple and effective techniques: initial residuals, Identity mapping and differentiable pooling. The hierarchical representation of the graph is realized. Each layer of the deep GNN learns the differentiable soft cluster allocation for the nodes, and maps the nodes to a set of clusters, and then these clusters are used as coarsening input and input to the next layer of GNN. Our experiments show that the deep VENUSAUR model outperforms the most advanced methods on various semi-supervised and fully-supervised tasks.