A Citation Recommendation Algorithm Based on SciBERT and a Temporal-Aware Graph Attention Network

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Citation recommendation plays a critical role in scholarly information retrieval by assisting researchers in identifying relevant and influential literature. Existing approaches typically rely on either textual semantic modeling or graph-based citation analysis, but often fail to jointly capture semantic relevance, structural dependencies, and temporal dynamics inherent in academic citation networks. In this paper, we propose SB-TGAT, a context-aware citation recommendation model that integrates scientific text representations with a Temporal-Aware graph attention mechanism. Specifically, we employ SciBERT to encode the semantic information of paper titles and abstracts, and design a Temporal-Aware Graph Attention Network (TGAT) to model citation relationships while explicitly incorporating temporal decay effects. The Temporal-Aware attention mechanism dynamically adjusts neighbor contributions according to publication time differences, enabling the model to balance long-term foundational works and short-term emerging research. To effectively integrate semantic and structural information, the textual embeddings and Temporal-Aware graph embeddings are fused through a feature interaction module and further optimized using a feed-forward neural network. We conduct extensive experiments on the DBLP-v14 dataset and evaluate the proposed model using standard information retrieval metrics, including MAP, MRR, and Recall@K. Experimental results demonstrate that SB-TGAT consistently outperforms strong baseline methods, validating the effectiveness of Temporal-Aware modeling and semantic– structural fusion for citation recommendation.