<![CDATA[Large language models (LLMs) reason well over natural language but do not naturally understand graph structure, and directly applying them to graph data runs into two persistent obstacles: a wide gap between the graph and text modalities, and the difficulty of encoding structural information in a way an LLM can use. Existing graph-instruction-tuning methods typically close this gap with simple linear projections and question-answer pairs that carry no explicit reasoning trace, which leaves cross-modal alignment shallow and drives up fine-tuning cost. This paper proposes GraphCoT, a graph representation learning method built on efficient chain-of-thought (CoT) fine-tuning. Its core is a high-quality CoT distillation mechanism in which a strong teacher model generates instruction data containing explicit intermediate reasoning paths, which explicitly teaches a student model to reason in multiple steps from only a small amount of training data. GraphCoT further introduces a graph-text alignment module that maps graph representations into the language embedding space, and a lightweight two-stage training strategy that aligns modalities before task-specific fine-tuning, which keeps training cost low while preserving both reasoning quality and transferability. Node classification and link prediction experiments on multiple benchmark datasets show that GraphCoT outperforms mainstream graph learning methods in accuracy and generalization, confirming the effectiveness of the CoT distillation mechanism and the alignment strategy for graph representation learning.]]>
Copyrights © 2024
