跳转至

文章背景与核心概要

本文为图神经网络（GNN）研究引入了一种全新的视角，将传统的消息传递机制重新构想为检索增强过程。尽管图神经网络通过邻域聚合在历史上取得了优于逐节点多层感知机（MLP）的性能，但其底层机制往往计算成本高昂，且容易受到结构噪声的影响。

为了克服这些局限性，作者提出了 RTA（Retrieval-augmented Text-attributed Graph Learning）这一精简的基于 MLP 的框架，用标签感知的检索与传播替代了结构化的消息传递。通过理论证明和在文本属性图上的广泛基准测试，RTA 能够匹敌甚至超越先进的 GNN 和图大模型（Graph LLM）基线，同时在面对检索错误时表现出更高的计算效率和鲁棒性。

Rethinking Message Passing as Retrieval for Text-Attributed Graph Learning

arXiv ID: arXiv:2608.26732 [cs.LG]
Submitted: August 27, 2026
Authors: Jintang Li, Yuhong Chen, Ruofan Wu, Binli Luo, Jiayi Ji, Hui Li, Rongrong Ji

arXiv ID: arXiv:2608.26732 [cs.LG]
Submitted: August 27, 2026
Authors: Jintang Li, Yuhong Chen, Ruofan Wu, Binli Luo, Jiayi Ji, Hui Li, Rongrong Ji

📌 Summary

本文从全新的视角审视了图神经网络（GNN），将传统的消息传递机制重新定义为一个检索增强的过程。

This paper introduces a fresh perspective on Graph Neural Networks (GNNs) by reformulating traditional message passing as a retrieval-augmented process.

尽管 GNN 历史上面向逐节点多层感知机（MLP）时通过邻域聚合取得了更好的效果，但其底层机制在计算上仍然较为昂贵，且容易受到结构噪声的干扰。为了解决这些局限性，作者提出了 RTA，这是一个简化的、基于 MLP 的框架，它用标签感知的检索与传播替代了结构化的消息传递。在理论证明和文本属性图的大规模基准测试支持下，RTA 能够匹配或超越先进的 GNN 和图大模型（Graph LLM）基线，同时具备更高的计算效率以及对检索错误异常值的鲁棒性。

While GNNs have historically outperformed node-wise Multi-Layer Perceptrons (MLPs) through neighborhood aggregation, the underlying mechanisms remain computationally expensive and vulnerable to structural noise. To address these limitations, the authors propose RTA, a streamlined, MLP-based framework that substitutes structural message passing with label-aware retrieval and propagation. Supported by theoretical proofs and extensive benchmarking on text-attributed graphs, RTA matches or exceeds state-of-the-art GNN and graph LLM baselines while offering greater computational efficiency and robustness against mis-retrieved outliers.

📝 Abstract

图神经网络（GNN）通常被概念化为消息传递神经网络，然而邻域 aggregation 为何能可靠地优于逐节点多层感知机（MLP）目前仍不十分清楚。尽管取得了实证成功，但这一范式可能带来高昂的计算成本，并且对不完美的图结构非常敏感。在这项工作中，我们提出了 GNN 的检索增强视角：每一层的预测都是通过将 MLP 应用于节点表示以及检索到的图上下文的置换不变摘要来实现的。基于这一视角，我们提出了 RTA，一个简单的基于 MLP 的框架，它用标签感知检索与传播替换了结构化的消息传递。我们提供了理论见解：（i）将基于检索的聚合与 softmax 注意力消息传递联系起来，以及（ii）确立了检索上下文监督对检索错误异常值的鲁棒性。在多个文本属性图基准上的实验表明，RTA 能够匹配甚至超越强大的 GNN 和图大模型基线，同时在各种多样化场景中提升了效率和鲁棒性。

Graph neural networks (GNNs) are typically conceptualized as message-passing neural networks, yet it remains unclear why neighborhood aggregation reliably outperforms node-wise multilayer perceptrons (MLPs). Despite its empirical success, this paradigm can be computationally expensive and sensitive to imperfect graph structures. In this work, we present a retrieval-augmented view of GNNs: each layer makes predictions by applying an MLP to a node representation together with a permutation-invariant summary of retrieved graph context. Motivated by this perspective, we propose RTA, a simple MLP-based framework that replaces structural message passing with label-aware retrieval and propagation. We provide theoretical insights that (i) connect retrieval-based aggregation to softmax-attention message passing, and (ii) establish the robustness of retrieved-context supervision to mis-retrieved outliers. Experiments on multiple text-attributed graph benchmarks show that RTA matches or even outperforms strong GNN and graph LLM baselines while improving efficiency and robustness across diverse scenarios.