PRISM: Reliable LLM Knowledge Distillation from Graph-Context Perspectives for Text-Attributed Graph Learning
Abstract
Text-attributed graph (TAG) learning aims to leverage both textual attributes and graph structures for node classification. Recent studies have utilized Large Language Models (LLMs) to generate pseudo-labels and rationales for few-label TAG learning. However, existing approaches typically rely solely on the textual attribute of target nodes, making the generated knowledge susceptible to noise and unreliability. To address this limitation, we propose PRISM (Perspective-based Reliable Imitation with Semantic Multiplicity), a reliability-aware knowledge distillation framework for few-label TAG learning. PRISM constructs multiple structural-context perspectives for each target node and estimates the reliability of LLM-generated knowledge based on cross-perspective consistency. The resulting reliability weights are then incorporated into the distillation process, enabling GNNs to place greater emphasis on more reliable supervision signals. Extensive experiments on five TAG benchmark datasets demonstrate that PRISM consistently outperforms existing methods in few-label node classification.