The content delves into the characterization of graph datasets for node classification by examining homophily measures and introducing label informativeness. The authors highlight the limitations of commonly used homophily measures and advocate for adjusted homophily as a more reliable alternative. Additionally, they propose label informativeness as a new characteristic to differentiate between different types of heterophilous graphs.
The discussion covers the properties desirable for a good homophily measure, such as maximal agreement, minimal agreement, constant baseline, empty class tolerance, and monotonicity. Adjusted homophily is presented as a superior measure that satisfies many of these properties compared to traditional measures like edge homophily and node homophily.
Furthermore, the concept of label informativeness is introduced to assess how much information a neighbor's label provides about a node's label. The authors demonstrate through experiments that label informativeness correlates better with Graph Neural Network (GNN) performance than traditional homophily measures.
The content also includes empirical illustrations using synthetic and semi-synthetic data to showcase the correlation between GNN performance and both homophily measures and label informativeness. Results indicate that label informativeness aligns more closely with GNN performance across various datasets.
Overall, the paper emphasizes the importance of considering adjusted homophily and label informativeness in characterizing graph connectivity patterns for node classification tasks.
In eine andere Sprache
aus dem Quellinhalt
arxiv.org
Wichtige Erkenntnisse aus
by Oleg Platono... um arxiv.org 03-05-2024
https://arxiv.org/pdf/2209.06177.pdfTiefere Fragen