HYDRA: Bridging Dynamic and Static User Modeling for Scientific Article Recommendation

In academic environments overwhelmed by the ever-growing volume of scientific publications, delivering personalized and adaptive recommendations remains a significant challenge, particularly due to the dynamic nature of user preferences. In this context, we developed a hybrid framework for scientific articles recommendation tailored to both static user profiles and evolving user behaviors. We leveraged a multi-source data pipeline—collecting from arXiv, Semantic Scholar, and PubMed—combined with thorough preprocessing and embedding generation to effectively represent article contents. Our hybrid framework integrates content-based filtering, collaborative filtering based on reading time, sequential modeling using LSTM (Long Short-Term Memory), and natural language explanations generated by a fine-tuned LLM (Large Language Model). The proposed framework offers three core contributions: (1) adaptability to contextual and temporal shifts in user preferences through the integration of sequential interactions; (2) robustness through the fusion of complementary recommendation techniques; and (3) interpretability via natural language explanations that enhance transparency and user trust. The architecture is designed with scalability in mind, making it suitable for deployment in real-world academic digital libraries. Rigorous experimentation was conducted using a controlled dataset of 1,000 synthetic users interacting with a curated subset of 10,000 articles, revealing the framework’s superior performance alongside notable diversity and adaptability to evolving user preferences.