From connectivity to insight: A graph neural network–large language model synergy for scholar–topic recommendation

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).

Research topic selection is a pivotal decision in a scholar’s academic trajectory, yet identifying viable topics remains challenging due to information overload within academic literature. To address these challenges, we propose a novel graph neural network–large language model synergy framework that systematically integrates structural connectivity learning with semantic expert reasoning for personalized scholar–topic recommendation. Unlike existing approaches that rely solely on graph topology or semantic generation, our framework decomposes the recommendation process into two stages: (1) constructing a scholar–paper–topic heterogeneous bibliometric network to capture long-term research trajectories through embedding learning with a heterogeneous graph neural network (HetGNN) and (2) empowering large language models as domain experts through chain-of-thought prompting to perform reasoning-based filtering and re-ranking. A key innovation of our approach lies in constraining large language models to reason over graph-derived candidate topics, which effectively mitigates the propensity for hallucinating irrelevant topics while addressing the “sparsity” issue inherent in purely graph-based methods. Empirical evaluation using Scopus publication data from scholars at Renmin University of China, along with cross-institutional validation (Peking University and Tsinghua University), demonstrates the framework’s superior performance. The synergy strategy significantly improves recommendation accuracy, increasing the F1-score up to 85.91%, and effectively bridges the gap between network topology and deep semantic insight in scientometric recommender systems.