RAR-SB: research article recommendation using SciBERT with BiGRU

The wide range and enormous volume of academic papers on the Internet prompted researchers to recommend models that could provide users with customized academic article recommendations. Nevertheless, previous approaches struggled with “sparsity” and “cold-start” as a consequence of a lack of sufficient information about research articles. Furthermore, they fail to recognize the importance of important factors and long-range dependencies, thus restricting their ability to make reliable and reasonable recommendations. To address these issues, we suggest RAR-SB, a research article recommender model that uses a pre-trained language model for scientific text named SciBERT to learn context-preserving research article representations. To learn the researcher’s preferences, the model exploits semantics corresponding to the title, abstract, authors, and field of study(FoS)/keywords of the candidate and query papers. The model captures long-range dependencies and salient features using the BiGRU network and the attention module, respectively. The experimental findings on the DBLP-V12 dataset demonstrate that the suggested recommendation model outperforms the baseline approaches regarding mean reciprocal rank (MRR) and mean average precision (MAP) by nearly 3.7% and 5.3%, respectively. Similarly, on the DBLP-V13 dataset, the proposed model has improved 6% and 5% better MRR and MAP results, respectively.