Personalized and diverse research topic recommendation for scholars: a graph convolutional network approach with large language model-based semantic enhancement

Personalized and diverse research topic recommendations are crucial for addressing academic information overload and promoting individual scholars’ innovation. To tackle the issues of data sparsity and semantic deficiencies, this paper proposes a novel topic recommendation model based on graph convolutional networks (GCN) and large language models (LLMs), called LLM-TRec. Specifically, the paper defines two types of topics for recommendation: newly explored topics (NETs) and historical memory topics (HMTs). A scholar-topic bipartite interaction graph (STIG) is constructed, and the lightweight LightGCN algorithm is used to extract topological features through multi-layer convolution. LLMs are employed to generate scholar profiles (SP) and topic summaries (TS) in short-text format. These texts are transformed into embeddings and dynamically fused with topological features using the gated linear unit (GLU) technique to achieve high-quality semantic enhancement. Finally, the paper explores the interpretability of the recommended topics by leveraging LLMs' ability for continuous dialogue. Using the PubMed Knowledge Graph (PKG) and American Physical Society (APS) as the data sources, the experimental results show that LLM-TRec outperforms 11 selected baseline models. For example, in terms of nDCG, it improves the best baseline performance by 9.43% to 11.79% in PKG-based experiment. Ablation studies indicate that both SP and TS are indispensable, with SP contributing more significantly. Hyperparameter investigation provides additional details. The recommendation performance for NETs and HMTs shows reasonable differences.