Hybrid Graph-Recurrent Architecture for Citation Recommendation via Future Embedding Forecasting

The rapid expansion of scientific literature has intensified the challenge of identifying relevant citations, particularly for newly published or under-cited papers. Traditional citation recommendation systems typically model static relationships or respond to past citation activity, offering limited predictive power for emerging works. In response, this thesis presents a temporal modeling framework for citation recommendation that anticipates future scholarly relevance by forecasting the latent representations of academic papers. Building on prior work that utilized Temporal Graph Networks (TGNs) to model dynamic citation flows, we propose Graph-Time, a hybrid architecture that integrates a Graph Transformer with a GRU-based time series predictor. The transformer captures structural and temporal dependencies within yearly citation snapshots, while the GRU forecasts a paper’s future embedding based on its past trajectory. To address cold-start cases, We inc We evaluate Graph-Time on the AMiner Citation Network V14, comprising over 5 million papers and 36 million citation links, using strict year-based splits to avoid information leakage. Compared to dynamic and static baselines, our model achieves substantial improvements in recommendation performance, including a 21 percent gain in MRR and notable gains in Precision@10 and Recall@10. Ablation studies confirm the contribution of each architectural component, and we discuss ethical considerations surrounding bias reinforcement in scholarly recommendations. This work contributes a scalable, future-aware framework for citation forecasting and lays a foundation for real-time, temporally sensitive academic recommender systems.