Heterophily and Graph Neural Networks: Past, Present and Future
Heterophily and Graph Neural Networks: Past, Present and Future
复制标题
DOI：
--
复制
发表时间：
2023
期刊：
影响因子：
--
通讯作者：
Jiong Zhu;Yujun Yan;Mark Heimann;Lingxiao Zhao;L. Akoglu;Danai Koutra
中科院分区：
文献类型：
--
作者：
Jiong Zhu;Yujun Yan;Mark Heimann;Lingxiao Zhao;L. Akoglu;Danai Koutra
最近，人们对理解图神经网络（gnn）在表现异质性的输入图上的性能很感兴趣，或者不同类别的节点倾向于连接。最初的研究结果表明，许多标准GNN模型在某些基准数据集上表现出高度的异质性，这促使人们对现有的和新的GNN设计进行研究，以改善这些背景下的学习。然而，进一步的分析表明，如果没有这些专门的设计，某些高度异恋的设置不会挑战gnn，这就提出了导致性能下降的真正因素的问题。在这项工作中，我们首先回顾了用于处理具有异质性的图的各种GNN设计，并检查了它们与其他GNN研究目标（如鲁棒性、公平性和过度平滑避免）的联系。接下来，我们进行了一项实证研究，探讨了gnn能够和不能有效执行的特定异恋图条件。我们的分析表明，尽管高异质性并不普遍阻碍传统GNN，但异质性图中的独特挑战，特别是与低度节点和复杂兼容模式的交织效应，需要专门针对异质性设计GNN。总之，我们讨论了未来的研究方向，旨在在更广泛的背景下推进对异源性对gnn影响的理解。
Recently, there has been interest in understanding the performance of Graph Neural Networks (GNNs) on input graphs exhibiting heterophily, or the tendency for nodes of different classes to connect. Initial findings showed that many standard GNN models struggled on certain benchmark datasets exhibiting high heterophily, prompting research into existing and novel GNN designs that improved learning in these contexts. However, further analyses revealed that certain highly heterophilous settings did not challenge GNNs without these specialized designs, raising questions about the true factors causing performance degradation. In this work, we first review various GNN designs proposed for handling graphs with heterophily, and examine their connections to other GNN research objectives such as robustness, fairness, and oversmoothing avoidance. Next, we conduct an empirical study to investigate the specific heterophilous graph conditions under which GNNs can and cannot perform effectively. Our analysis reveals that although high heterophily does not universally impede conventional GNNs, unique challenges in heterophilous graphs, particularly the intertwined effects with low-degree nodes and complex compatibility patterns, warrant GNN designs specifically tailored to heterophily. In conclusion, we discuss future research directions aimed at advancing the understanding of the impact of heterophily on GNNs across a broader range of contexts.