Requirements tracing plays an important role to manage requirements and its related artifacts through the entire software life cycle. As manually creating such trace links is time-consuming and error-prone, some information retrieval (IR) based and machine learning (ML) based solutions have been proposed. Among them, unsupervised ML methods which do not require large labeled datasets are gaining more attention. Most of these solutions model the lexical and semantic information to resolve the problem. However, we find that existing approaches typically neglect the word co-occurrence distribution and word order information of the textual artifacts, which could provide extra indications for enhancing trace links. In this paper, we propose a novel approach, named GeT2Trace, which utilizes a graph mining-based expansion learning to enhance trace links recovery. The key idea is to exploit the word co-occurrence information and the word order information via graph network, and leverage them to learn a more comprehensive and accurate artifact representation. Evaluation is conducted on five public datasets, and the results show that our approach outperforms the state-of-the-art baselines. Expanding requirements with graphic information provide new insights into the unsupervised traceability solutions, and the improved trace links confirm the usefulness and effectiveness of GeT2Trace.