New Framework Rethinks Graph Neural Networks as Retrieval

Graph Neural Networks (GNNs) are typically understood through the lens of message passing, where information is aggregated from neighboring nodes. However, the exact reasons why this aggregation consistently outperforms simpler node-wise Multi-Layer Perceptrons (MLPs) remain somewhat unclear, and message passing can be computationally intensive and sensitive to imperfect graph structures. This research introduces a new perspective, viewing GNNs as a retrieval-augmented process. Under this new view, each layer of a GNN makes predictions by applying an MLP to a node's representation combined with a permutation-invariant summary of retrieved graph context. Motivated by this, the authors propose RTA (Retrieval-augmented Text-attributed graph learning), a straightforward MLP-based framework. RTA replaces the conventional structural message passing with a label-aware retrieval and propagation mechanism. Theoretical insights provided by the study connect retrieval-based aggregation to softmax-attention message passing and establish the robustness of retrieved-context supervision against mis-retrieved outliers. Experimental results on several text-attributed graph benchmarks demonstrate that RTA performs comparably to, or even surpasses, strong GNN and graph LLM baselines. Crucially, it achieves these results with improved efficiency and robustness across various scenarios.