提出异步消息传递机制，缓解图神经网络中远距离信息传播受阻问题。
Asynchronous Message Passing for Addressing Oversquashing in Graph Neural Networks
- 按中心性分批异步更新节点特征，避免信息集中压缩。
- 在REDDIT-BINARY和Peptides-struct上分别提升5%和4%准确率。
- 无需修改模型结构，适用于多种图分类任务。
图神经网络存在过度压缩问题，即结构瓶颈限制了远距离节点间的信息传递，影响需要长程交互的任务。现有方法受限：图重连会破坏归纳偏置，增加通道容量则引入更多参数。本文提出一种高效、模型无关的异步消息传递框架，各层节点特征异步更新，而非同步进行。每层仅由中心性引导的一批节点更新，使信息以串行方式传播，而非同时压缩至固定容量通道。理论证明，该框架的敏感度界随深度衰减更慢。在六个标准与两个长程图分类基准上应用，取得显著性能提升，其中REDDIT-BINARY和Peptides-struct分别提升5%和4%。
原文摘要 · Abstract (English)
Graph Neural Networks (GNNs) suffer from oversquashing, where structural bottlenecks limit message propagation between distant nodes, hindering tasks that require long-range interactions. Existing remedies are limited: graph rewiring alters edge connectivity, compromising inductive bias, while increasing channel capacity adds parameters. In this work, we propose an efficient, model-agnostic framework that asynchronously updates node features across layers, unlike standard synchronous message passing. At each layer, only a centrality-guided batch of nodes updates, letting information propagate sequentially rather than compress simultaneously into fixed-capacity channels. We show theoretically that our framework's sensitivity bound decays more slowly with depth than synchronous message passing. We have applied the framework to six standard and two long-range graph classification benchmarks, and it achieves notable gains, including $5\%$ and $4\%$ improvements on REDDIT-BINARY and Peptides-struct, respectively.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。