Bulbasaur–One Stitched Toy Graph Neural Network

Shanzai CL

Published on: 2021-08-12

Abstract

Graph neural networks have shown significant success in the field of graph representation learning. Graph convolution performs neighborhood aggregation and represents one of the most important graph operations. However, one layer of these neighbor aggregation methods only considers direct neighbors, and when several recent studies attributed this performance degradation to an over-refined problem, the problem pointed out that repeated propagation makes it difficult to distinguish between different types of representations. In this work, we believe that the key factor affecting performance is the entanglement of representation conversion and propagation in current graphics convolution operations. After decomposing these two operations, a deeper graph neural network can be used to learn the graph level representation from the above receptive domain. three simple and effective techniques are used in combination:differentiable pooling initial residuals and Identity mapping. 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.Based on our theoretical and empirical analysis, we propose a Deep Adaptive Initial Residual Identity Mapping Graph Neural Network (BULBASAUR) to adaptively merge information from large receiving fields.