Beyond Layers: A Global Message-Passing Mechanism for Heterophilic Graphs
Keywords: Message Passing Mechanism; Graph Neural Network; Heterophilic Graph
TL;DR: A global message-passing graph neural network to address the limitation of the localized layer-by-layer property of the message-passing mechanism in heterophilic graph.
Abstract: The effectiveness of most graph neural networks is largely attributed to the message-passing mechanism.
Despite the significant success in homophilic graphs (i.e., similar nodes are connected by edges), message-passing mechanism in heterophilic graphs (i.e., dissimilar nodes are connected by edges) is still challenging.
Due to the existence of low-order but dissimilar neighbor nodes in a path, messages from similar but high-order neighbor nodes are often weakened.
In this paper, firstly, we conduct both theoretical and empirical analysis of the layer-by-layer local nature of the message-passing mechanism.
Then, we propose a novel GloMP-GNN for heterophilic graphs by comprehensively introducing global insights into the message-passing mechanism.1) During the message propagation phase, the global insight is introduced from the perspective of graph structure.
We design a structure-based global propagation strategy, where messages can be effectively propagated with the bridge of virtual edges between a global virtual node and graph nodes.
Moreover, a global edge adaption approach is included to aggregate messages with adaptive edge weight adjustment.
2) During the feature updating phase, the global insight is introduced with a feature-augmented compensatory updating method.
Through a multi-view feature updating mechanism, the node feature representation can be effectively augmented by compensating the weakened message from different views.
Finally, we conduct extensive experimental evaluations on eight datasets, which demonstrate the superiority of our proposed GloMP-GNN. As broader impacts, GloMP-GNN consistently performs well across multiple layers and also effectively prevents the over-smoothing problem.
Codes are available on Github with https://github.com/Anonymous-GloMP-GNN/GloMP-GNN.
Primary Area: learning on graphs and other geometries & topologies
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 13497
Loading