Hybrid Citation Recommendation Utilizing Siamese BERT and Dynamic Graph Attention Network

In response to the existing limitations of citation recommendation algorithms, which primarily focus on citation relationships and lack in-depth analysis of citation content, as well as the utilization of entities and relationships within the knowledge graph, this paper put forwards a hybrid citation recommendation algorithm called HC-SBDGA (Hybrid Citation recommendation utilizing Siamese BERT and dynamic graph attention networks). Experimental verification was carried to verify the performance of the algorithm using public datasets such as DBLP [1] (Database Systems and Logic Programming) and AAN [2] (ACL Anthology Network). From the view of experiment results, an improved model contrasted to the baseline model is introduced. Specifically, the AUC value of the proposed model increased by 4.5 and 3.9 percentage points, and the F1 value increased by 5.6 and 3.7 percentage points. Precision@K and recall@K also showed varying degrees of improvement for different values of k. These results confirm that HC-SBDGA possesses strong recommendation capability, thereby better fulfilling the needs of researchers.