Amazon Computers, Amazon Photo, Coauthor CS, Coauthor Physics, Cora, CiteSeer, PubMed
收藏资源简介:
本文研究了七个常用的图学习数据集,包括Amazon Computers、Amazon Photo、Coauthor CS、Coauthor Physics、Cora、CiteSeer和PubMed。这些数据集主要用于图神经网络的基准测试,涵盖了社交网络、合作网络和引文网络等多种类型。数据集大小适中,具有小世界现象,适合进行图结构分析。研究通过对比MLP和图神经网络在这些数据集上的表现,发现MLP在某些数据集上表现出色,甚至超过了图神经网络,表明这些数据集的特征已经包含了足够的图信息,使得图结构的使用变得不那么必要。研究还提出了新的合成数据集,旨在更好地评估图神经网络的性能。
This paper investigates seven widely-used graph learning datasets, including Amazon Computers, Amazon Photo, Coauthor CS, Coauthor Physics, Cora, CiteSeer, and PubMed. These datasets are primarily employed for benchmarking graph neural networks (GNNs), covering diverse types such as social networks, co-authorship networks, and citation networks. With moderate scale and the small-world phenomenon, these datasets are well-suited for graph structure analysis. By comparing the performance of Multi-Layer Perceptrons (MLPs) and GNNs on these datasets, this study finds that MLPs deliver outstanding performance on certain datasets and even outperform GNNs, indicating that the node features of these datasets already encompass sufficient graph-related information, making the utilization of graph structures less necessary. Additionally, this paper proposes a novel synthetic dataset intended to better evaluate the performance of graph neural networks.




