Addressing the challenges of coexisting objects such as books, journal articles, theses, and course resources in university digital libraries, as well as weak cross-database associations and insufficient hierarchical organization, this paper proposes a cross-domain knowledge item association retrieval method based on the Graphormer-Lite graph encoding model. This research firstly builds a heterogeneous graph that includes books, journal papers, dissertations, course materials, subject words, writers, universities/branches of learning, and database origins, thus it unifies textual content, catalog properties, structure connections, and user actions in one single expression space. Based upon this foundation, the research puts forward relationship type bias, path distance coding, and local candidate attention, therefore it reduces the inference extra cost via layer compression, parameter sharing, and distillation methods. Experiments were conducted using 6 subject domains, 120,000 core knowledge entry nodes, approximately 2.15 million relationship edges, and 38,624 click and download logs. Results show that Graphormer-Lite achieves 0.684, 0.781, 0.643, and 0.612 on Recall@10,Recall@20, nDCG@10, and MRR of 0.684, 0.781, 0.643, and 0.612, respectively, outperforming BM25, Sentence-BERT, HAN, HGT, and the standard Graphormer overall; it also maintains a stable advantage in long-tail and cold-start scenarios while achieving a balance between latency and parameter size that is more suitable for online deployment. This method provides a reference for optimizing cross-domain resource discovery and association services in university digital libraries.
