基于PyG库的多种类图神经网络对比研究

A Comparative Study of Multi-class Graph Neural Networks Base on PyG Library

各种各样的基于消息传递机制的图神经网络架构被提出,但是这些图神经网络在同一节点分类任务的表现性能如何缺少相应的研究,文章基于PyG图神经网络库,编写实验代码,包括数据集处理、图神经网络搭建和节点分类模型训练3个主要环节,实现对不同架构的图神经网络在节点分类任务上的对比,通过利用多指标衡量基于多种类的图神经网络构建的节点分类模型,最后得出最适合节点分类的图神经网络结构,并给出实验分析,同时提出了今后的研究方向。

Various graph neural network architectures based on message passing mechanisms have been proposed,but there is a lack of corresponding research on how these graph neural networks perform on the same node classification task.Based on the PyG graph neural network library,our paper writes experimental code,including data set processing,graph neural network construction and node classification model training.We realize the comparison of graph neural networks with different architectures on the node classification task,and measure the node classification model based on multiple types of graph neural networks by using multiple metrics.Finally,the graph neural network structure that is most suitable for node classification is obtained,the experimental analysis is given,and the future research direction is also proposed.