A comparative evaluation of biomedical similar article recommendation

Area of Science:

• Biomedical Sciences
• Information Retrieval
• Computational Biology

Background:

• The rapid growth of biomedical literature poses challenges for researchers seeking relevant information.
• Automated article recommendation systems are crucial for efficient discovery of scientific findings.
• A comprehensive, algorithm-level comparison of existing recommendation methods is lacking.

Purpose of the Study:

• To investigate and empirically evaluate 15 diverse automated article recommendation methods for the biomedical domain.
• To compare the performance of term-based, embedding-based, and BERT-based approaches.
• To assess recommendation strategies in both article-oriented and user-oriented scenarios.

Main Methods:

• Evaluation of 15 distinct article recommendation algorithms.
• Categorization of methods into term-based, word embedding, sentence embedding, document embedding, and BERT-based approaches.
• Testing across two benchmark datasets (TREC 2005 Genomics, RELISH) in article- and user-oriented settings.

Main Results:

• BERT and BioSenVec text representation models demonstrated superior performance compared to traditional methods (e.g., BM25) and existing systems.
• These advanced models outperformed others across multiple evaluation metrics on both datasets.
• Fine-tuning BERT-based methods further enhanced their recommendation accuracy.

Conclusions:

• The study provides valuable insights for selecting optimal modeling strategies for biomedical article recommendation systems.
• BERT and BioSenVec represent promising approaches for enhancing information retrieval in biomedical research.
• Publicly available code and datasets facilitate reproducibility and further research.