ReTAG: A Retrieved Cellular Topologies-Augmented Graph Learning Framework

Sen Zhao, Jia Tang, Yuqi Sun, Qinghua Zhang, Xu Zhang

Keywords: Graph Neural Networks, Retrieval-Augmented Generation, Cell Complexes, Graph Prompt Tuning

Abstract: Retrieval-augmented graph learning (RAG) enhances the generalization of Graph Neural Networks (GNNs) by retrieving and integrating structurally relevant subgraphs, addressing their limitations on unseen or distribution-shifted graphs. However, current RAG-based methods mainly operate on zero-(nodes) and one-dimensional (edges) elements, failing to capture higher-dimensional topological structures, such as cycles, that are essential for identifying critical substructures and modeling complex relational patterns. This limitation hinders the retrieval of high-dimensional topological characteristics and weakens reasoning over graphs with complex higher-order interactions. In this paper, we propose a novel Retrieved Cellular Topologies-Augmented Graph Learning Framework (ReTAG), that leverages cellular complexes to model and retrieve multi-dimensional topology-aware subgraphs, termed cellular topologies. These structures encode multi-dimensional topological interactions across nodes, edges, and higher-dimensional cells. During inference, ReTAG retrieves cellular topologies based on their topological and semantic alignment with the input graph, and integrates them via a multi-dimensional topological message-passing mechanism that enables effective propagation of topological information across dimensions. Experiments on node classification, link prediction, and graph classification show ReTAG outperforms existing methods. The implementation code is available in the supplementary material.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 6000