Learning to rank figures within a biomedical article.

Area of Science:

• Biomedical Informatics
• Artificial Intelligence in Science
• Scientific Literature Analysis

Background:

• Biomedical literature contains millions of figures crucial for research validation and hypothesis generation.
• Current figure search tools are limited by the "bag of figures" approach, neglecting inter-figure relationships.
• Efficiently accessing relevant figures is challenging due to the sheer volume of published research.

Purpose of the Study:

• To develop and evaluate computational approaches for automatically ranking the importance of figures in biomedical articles.
• To introduce BioFigRank, a novel supervised-learning model for figure ranking based on expert annotations.
• To compare the performance of BioFigRank against state-of-the-art models and human experts.

Main Methods:

• Leveraged and extended state-of-the-art listwise learning-to-rank algorithms.
• Developed a supervised-learning model, BioFigRank, using a dataset of researcher-annotated figure importance.
• Employed greedy feature selection to optimize model performance.
• Evaluated BioFigRank against First Authors, Non-Author-In-Domain-Experts, and Non-Author-Out-Domain-Experts.

Main Results:

• BioFigRank demonstrated superior performance compared to other computational models in cross-validation.
• Greedy feature selection significantly enhanced BioFigRank's ranking performance.
• BioFigRank outperformed Non-Author-Out-Domain-Experts and performed comparably to Non-Author-In-Domain-Experts.
• While underperforming First Authors, BioFigRank achieved expert-level performance for most user scenarios.

Conclusions:

• BioFigRank offers an effective AI-driven solution for navigating and prioritizing figures in biomedical literature.
• The model provides expert-level intelligence, significantly improving efficiency for researchers.
• BioFigRank addresses the limitations of current search methods by considering figure relationships and importance.