Kernkonzepte
High-degree nodes tend to have a lower probability of misclassification in graph neural networks, due to factors associated with node degree such as neighborhood homophily and diversity.
Zusammenfassung
The paper provides a theoretical and empirical analysis of the origins of degree bias in graph neural networks (GNNs).
Key highlights:
- The authors survey prior hypotheses for degree bias and find they are often not rigorously validated or can be contradictory.
- They prove that high-degree test nodes tend to have a lower probability of misclassification, regardless of how GNNs are trained. This is due to factors associated with node degree, such as neighborhood homophily and diversity.
- For the random walk (RW) GNN, the authors show that low-degree nodes tend to have higher variance in their representations, leading to a higher probability of being misclassified.
- For the symmetric normalized (SYM) GNN, the authors show that the loss on low-degree nodes may be adjusted more slowly during training compared to high-degree nodes.
- Despite these training discrepancies, the authors empirically demonstrate that message-passing GNNs can achieve their maximum possible training accuracy, which is not significantly limited by their expressive power.
- Based on their theoretical and empirical insights, the authors provide a roadmap to alleviate degree bias in GNNs.
Statistiken
The paper does not contain any key metrics or important figures to support the author's key logics.
Zitate
The paper does not contain any striking quotes supporting the author's key logics.