A paper recommendation method based on SPECTER and graph convolutional networks

To address the issues of data sparsity and cold-start problems in paper recommendation, as well as the limitations of general document representation methods in representing paper datasets, this study proposes a paper recommendation method based on SPECTER and Graph Convolutional Networks (PR-SGCN method). This method leverages pre-trained document-level representation learning and the citation-aware transformer SPECTER to learn paper content representations, thereby overcoming the limitations of existing methods in representational capability. Additionally, it employs network representation learning and Graph Convolutional Networks (GCN) to mine hidden information from both attribute and structural perspectives, effectively mitigating the problems of data sparsity and cold starts. Experiments conducted on the ACL Anthology Network (AAN) and DBLP datasets demonstrate that, compared with the academic paper recommendation method based on heterogeneous graphsCitation Recommendation based on Weighted Heterogeneous Information Network with Citation Semantic Links (WHIN-CSL)and the academic literature recommendation method integrating network representation learning and textual information, the PR-SGCN method achieves improvements in Recall@25 by 59 and 72 percentage points, and by 67 and 65 percentage points, respectively.