Why was this cited? Explainable machine learning applied to COVID-19 research literature

Area of Science:

• Bibliometrics and Scientometrics
• Computational Biology
• Machine Learning in Scientific Research

Background:

• Traditional bibliometric factors are insufficient for predicting research citation counts.
• The COVID-19 crisis highlighted the need for rapid analysis of vast biomedical literature.
• Understanding citation patterns can guide research funding and identify impactful studies.

Purpose of the Study:

• To develop machine learning models predicting citation counts using article content and metadata.
• To identify novel patterns and factors influencing research impact beyond traditional bibliometrics.
• To compare the performance and interpretability of various machine learning and rule-learning algorithms.

Main Methods:

• Utilized the CORD-19 corpus comprising biology and medicine research articles.
• Employed advanced machine learning for text understanding: BERT, ConceptNet, Pubtator, ScispaCy.
• Applied explanation algorithms (Random Forest, LIME, Shapley values) and compared black-box models with intrinsically explainable rule-learning models (CORELS, CBA).

Main Results:

• Identified significant patterns, including associations with dipeptidyl peptidase 4 (DPP4), a MERS-CoV receptor.
• Found that articles referencing bats and camels as hosts for coronaviruses received more citations.
• Observed potential citation bias favoring authors with Western-sounding names; TF-IDF and binary word incidence showed equal performance, with the latter offering better interpretability.

Conclusions:

• Machine learning, particularly neural networks, can effectively predict citation counts using article content and metadata.
• Rule-based models, enhanced by semantic entity detection, provide valuable insights into citation drivers.
• Further research should explore citation patterns in phylogenetic contexts and investigate DPP4's role in SARS-CoV-2.