Monophilic Neighbourhood Transformers
Keywords: graph neural networks, transformers
TL;DR: A novel message exchanging mechanism is implemented by applying Transformers in every neighbourhood, achieving SOTA on homophilic and heterophilic graphs.
Abstract: Graph neural networks (GNNs) have seen widespread application across diverse fields, including social network analysis, chemical research, and computer vision.
Nevertheless, their efficacy is compromised by an inherent reliance on the homophily assumption, which posits that adjacent nodes should exhibit relevance or similarity.
This assumption becomes a limitation when dealing with heterophilic graphs, where it is more common for dissimilar nodes to be connected.
Addressing this challenge, recent research indicates that real-world graphs generally exhibit monophily, a characteristic where a node tends to be related to the neighbours of its neighbours.
Inspired by this insight, we introduce Neighbourhood Transformers (NT), a novel approach that employs self-attention within every neighbourhood of the graph to generate informative messages for the nodes within, as opposed to the central node in conventional GNN frameworks.
We develop a neighbourhood partitioning strategy equipped with switchable attentions, significantly reducing space consumption by over 95\% and time consumption by up to 92.67\% in NT.
Experimental results on node classification tasks across 5 heterophilic and 5 homophilic graphs demonstrate that NT outperforms current state-of-the-art methods, showcasing their expressiveness and adaptability to different graph types.
The code for this study is available at https://anonymous.4open.science/r/MoNT-BD3C .
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: 10389
Loading