Beyond Homophily in Graph Neural Networks: Current Limitations and Effective Designs

By Jiong Zhu, Yujun Yan, Lingxiao Zhao, Mark Heimann, Leman Akoglu, Danai Koutra. Neural Information Processing Systems, 2020

1,400 citations recorded by the source corpus.

View Paper: https://www.semanticscholar.org/paper/21e33bd0ad95ee1f79d8b778e693fd316cbb72d4

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

1. Xiao Wang
2. Oleg Platonov
3. Yao Ma
4. Joost van de Weijer
5. Jiliang Tang
6. Artem Babenko
7. Sitao Luan
8. Chuan Shi

Citations in the corpus

• Exploring the Potential of Large Language Models (LLMs)in Learning on Graphs (2023, background): “By following the "homophily" assumption [54, 55], which suggests that neighboring nodes tend to share the same labels, the inclusion of neighboring information can potentially alleviate annotation bias.”
• A Survey on Oversmoothing in Graph Neural Networks (2023, unknown): “…being a key limitation in the development of deep multi-layer GNNs, over-smoothing can also severely impact the ability of GNNs to handle heterophilic graphs (Zhu et al., 2020), in which node labels tend to differ from the labels of the neighbors and thus long-term interactions have to be learned.”
• A Comprehensive Survey on Community Detection With Deep Learning (2021, background): “Network heterophily [179] can be interpreted as a phenomenon, where connected nodes belong to different communities or are characterized by dissimilar features.”
• Graph Neural Networks with Heterophily (2020, methodology): “…under heterophily (Zhu et al. 2020): GCN (Kipf and Welling 2017), GAT (Veli ˇ ckovi ´ c et al. 2018), GCN-Cheby (Deffer-rard, Bresson, and Vandergheynst 2016; Kipf and Welling 2017), GraphSAGE (Hamilton, Ying, and Leskovec 2017), MixHop (Abu-El-Haija et al. 2019), and H 2 GCN (Zhu et al. 2020).”
• Finding Global Homophily in Graph Neural Networks When Meeting Heterophily (2022, methodology): “As a representative model, H2GCN (Zhu et al., 2020b) presents three designs to improve the performance of GNNs under heterophily: ego and neighbor embedding separation, higher-order neighborhood utilization and intermediate representation combination.”
• Exploiting the Intrinsic Neighborhood Structure for Source-free Domain Adaptation (2021, background): “Furthermore, inspired by recent graph based methods [1, 3, 61] which show that the higher order neighbors can provide relevant context, and also considering neighbors of neighbors is more likely to provide datapoints that are close on the data manifold [43].”
• A critical look at the evaluation of GNNs under heterophily: are we really making progress? (2023, methodology): “Our results also show that there is, however, a trick that is useful for learning on heterophilous graphs — separating ego-and neighbor-embeddings, which was proposed in Zhu et al. (2020).”
• Revisiting Heterophily For Graph Neural Networks (2022, methodology): “In some cases, even simple Multi-Layer Perceptrons (MLPs) can outperform GNNs by a large margin on relational data [46, 29, 32, 8].”
• Is Homophily a Necessity for Graph Neural Networks? (2021, methodology): “Due to this neighborhood aggregation mechanism, existing literature posits that strong homophily of the underlying graph is a necessity for GNNs to achieve good performance on SSNC [1, 41, 40, 4, 22, 12, 14, 35].”
• Beyond Low-frequency Information in Graph Convolutional Networks (2021, background): “This is a strange phenomenon, because MLP uses raw features as input, which contains high-frequency information, so its performance should be better than GCN (Zhu et al. 2020).”