Structure-Aware Multi-Hop Graph Convolution for Graph Neural Networks

Yang Li,Yuichi Tanaka

doi:10.1109/access.2022.3149619

Abstract

We present a spatial graph convolution (GC) to classify signals on a graph. Existing GC methods are limited in using the structural information in the feature space. Furthermore, GCs only aggregate features from one-hop neighboring nodes to the target node in their single step. In this paper, we propose two methods to improve the performance of GCs: 1) Utilizing structural information in the feature space, and 2) exploiting the multi-hop information in one GC step. In the first method, we define three structural features in the feature space: feature angle, feature distance, and relational embedding. The second method aggregates the node-wise features of multi-hop neighbors in a GC. Both methods can be simultaneously used. We also propose graph neural networks (GNNs) integrating the proposed GC for classifying nodes in 3D point clouds and citation networks. In experiments, the proposed GNNs exhibited a higher classification accuracy than existing methods.

Highlights

D EEP neural networks have gained remarkable success in detecting, segmenting, and recognizing regularly structured data such as images and videos [1]–[3], owing to large-scale datasets and advancements in computing power
We consider there exist two limitations: 1) Existing methods might lose the structural information of the surrounding neighboring nodes, in the feature space
STRUCTURAL FEATURES We present three structural features in the feature space to obtain precise structural information of the surrounding onehop neighboring nodes:

Summary

Introduction

D EEP neural networks have gained remarkable success in detecting, segmenting, and recognizing regularly structured data such as images and videos [1]–[3], owing to large-scale datasets and advancements in computing power. Many irregular data without a fixed sample order exist in the real world, which are challenging to process using classic deep learning approaches. Such examples include attributes on 3D point clouds, opinions on social networks, and the number of passengers in traffic networks. Structured data could be processed more effectively by transforming those to graph-structured data In this context, graph neural networks (GNNs) have received a lot of attention [4]– [6]. In the single step of GCs, existing methods primarily focus on using node-wise features of the one-hop neighborhood [9]–[11]. We consider there exist two limitations: 1) Existing methods might lose the structural information of the surrounding neighboring nodes, in the feature space. 2) While the multi-hop neighborhood may contain some useful information, this cannot be utilized for the single step GC

Objectives

Results

Conclusion