LLMTGL: Large Language Models for Robust Text-Attributed Graph Learning under Missing Text and Label Noise
Abstract
Text-attributed graphs have attracted considerable attention due to their extensive applications in domains such as social networks and citation networks. Existing methods typically rely on reliable text-attributed graphs. However, the frequent occurrence of missing text in real-world scenarios severely constrains model performance. Meanwhile, label noise is inevitably introduced during the data annotation process, leading to a degradation in the performance of graph learning. The large language model (LLM), with their text generation and semantic understanding, offers a unified solution to the challenges of missing text and label noise. To jointly address both challenges, we propose LLMTGL, a novel graph learning framework for text-attributed graph node classification. Specifically, we introduce graph topological properties to construct augmented prompts that guide the LLM in generating semantically coherent academic text descriptions for nodes with missing text. Subsequently, we design an LLM-driven pseudo-label expansion mechanism that selects high-confidence unlabeled nodes to the training set. This framework jointly optimizes for missing text and label noise by integrating graph topological structure with semantic information. Experimental results demonstrate that LLMTGL achieves superior performance compared to state-of-the-art methods on text-missing graphs with noisy labels. The source code is available at https://anonymous.4open.science/r/LLMTGL-EC24ssjs.