Motif-Aware Embeddings for Enhanced Few-Shot Graph Learning

ICLR 2026 Conference Submission21928 Authors

Keywords: Graph few shot learning, Motif

TL;DR: A novel framework that integrates motif representations into graph few-shot learning model

Abstract: Graph Neural Networks (GNNs) have shown strong performance in node classification tasks. However, in real-world scenarios, only a limited number of nodes are often labeled, leading to the few-shot node classification problem, which is a significant challenge for GNNs. Most existing research focuses on designing new models to adapt to this setting, but often overlooks structural information in the graph data, such as motif patterns, which can provide crucial cues for learning from a few examples. In this paper, we propose a novel framework that integrates motif representations into graph few-shot learning models. Specifically, we extract unique motif representations from the graph and introduce them as virtual nodes. To capture richer structural patterns, we further enhance motif extraction by adding cluster labels based on node similarity, thereby incorporating both structural and feature information. Additionally, we assign TF-IDF scores as edge weights between virtual motif nodes and original nodes to quantify the importance of their connections. Experimental results demonstrate that our approach consistently improves the performance of various graph few-shot learning methods.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 21928