This website requires JavaScript.

2-hop Neighbor Class Similarity (2NCS): A graph structural metric indicative of graph neural network performance

Andrea CavalloClaas GrohnfeldtMichele RussoGiulio LovisottoLuca Vassio
Dec 2022
摘要
Graph Neural Networks (GNNs) achieve state-of-the-art performance ongraph-structured data across numerous domains. Their underlying ability torepresent nodes as summaries of their vicinities has proven effective forhomophilous graphs in particular, in which same-type nodes tend to connect. Onheterophilous graphs, in which different-type nodes are likely connected, GNNsperform less consistently, as neighborhood information might be lessrepresentative or even misleading. On the other hand, GNN performance is notinferior on all heterophilous graphs, and there is a lack of understanding ofwhat other graph properties affect GNN performance. In this work, we highlight the limitations of the widely used homophily ratioand the recent Cross-Class Neighborhood Similarity (CCNS) metric in estimatingGNN performance. To overcome these limitations, we introduce 2-hop NeighborClass Similarity (2NCS), a new quantitative graph structural property thatcorrelates with GNN performance more strongly and consistently than alternativemetrics. 2NCS considers two-hop neighborhoods as a theoretically derivedconsequence of the two-step label propagation process governing GCN'straining-inference process. Experiments on one synthetic and eight real-worldgraph datasets confirm consistent improvements over existing metrics inestimating the accuracy of GCN- and GAT-based architectures on the nodeclassification task.
展开全部
图表提取

暂无人提供速读十问回答

论文十问由沈向洋博士提出,鼓励大家带着这十个问题去阅读论文,用有用的信息构建认知模型。写出自己的十问回答,还有机会在当前页面展示哦。

Q1论文试图解决什么问题?
Q2这是否是一个新的问题?
Q3这篇文章要验证一个什么科学假设?
0
被引用
笔记
问答