A Hybrid Citation Recommendation Model With SciBERT and GraphSAGE

As the number of scientific publications continues to increase at a dizzying rate, researchers face challenges related to spending too much time and effort searching for appropriate papers to cite in their work. Citation recommendation models have thus been developed to automatically generate a list of relevant papers for a specific text passage, thus helping to reduce the workload for scientists and contribute to better-quality research. Consequently, this research direction has recently attracted significant interest in the scientific community. However, the current citation recommendation models still focus primarily on the citation context and do not adequately address the metadata of papers, such as the citation links, publication time, and venue. To overcome these problems, in this study, we propose the SciBERT-GraphSAGE which is a hybrid deep learning-based model for recommending a list of academic papers by considering both the citation context and this article's metadata. Our model has two important components: 1) SciBERT for text data representation learning and 2) GraphSAGE for learning the representations of this article's citation links. We validate the effectiveness of our model on three benchmark datasets: 1) FullTextPeerRead; 2) ACL; and 3) RefSeer. The results from experiments demonstrate that our novel SciBERT-GraphSAGE model outperforms previous advanced models in terms of Recall@K, mean reciprocal rank (MRR), and mean average precision (MAP).