Comparing Text-Based Similarity Approaches for Clustering Biomedical Publications
Author Information
Author(s): Kevin W. Boyack, David Newman, Russell J. Duhon, Richard Klavans, Michael Patek, Joseph R. Biberstine, Bob Schijvenaars, André Skupin, Nianli Ma, Katy Börner
Primary Institution: SciTech Strategies, Inc.
Hypothesis
Which similarity approach generates the most coherent clusters of a biomedical literature set of over two million documents?
Conclusion
PubMed's related article approach generated the most coherent clusters, followed closely by the BM25 approach using titles and abstracts.
Supporting Evidence
- The study analyzed over two million biomedical documents to assess clustering accuracy.
- Nine different similarity approaches were tested for their effectiveness in clustering.
- PMRA and BM25 approaches outperformed others in generating coherent clusters.
- The study provides a large-scale comparison of text-based similarity methods.
Takeaway
The study looked at different ways to group similar medical articles and found that one method worked best for making clear groups.
Methodology
The study used a corpus of 2.15 million records from MEDLINE and generated nine different document-document similarity matrices from bibliographic records.
Potential Biases
The validation inputs (titles and abstracts) may not be independent of the clustering inputs, potentially introducing bias.
Limitations
The study only focused on text-based similarity approaches and did not explore all possible methods for clustering.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website