Clustering more than two million biomedical publications: comparing the accuracies of nine text-based similarity approaches

Abstract

Background: We investigate the accuracy of different similarity approaches for clustering over two million biomedical documents. Clustering large sets of text documents is important for a variety of information needs and applications such as collection management and navigation, summary and analysis. The few comparisons of clustering results from different similarity approaches have focused on small literature sets and have given conflicting results. Our study was designed to seek a robust answer to the question of which similarity approach would generate the most coherent clusters of a biomedical literature set of over two million documents.

Methodology: We used a corpus of 2.15 million recent (2004-2008) records from MEDLINE, and generated nine different document-document similarity matrices from information extracted from their bibliographic records, including titles, abstracts and subject headings. The nine approaches were comprised of five different analytical techniques with two data sources. The five analytical techniques are cosine similarity using term frequency-inverse document frequency vectors (tf-idf cosine), latent semantic analysis (LSA), topic modeling, and two Poisson-based language models--BM25 and PMRA (PubMed Related Articles). The two data sources were a) MeSH subject headings, and b) words from titles and abstracts. Each similarity matrix was filtered to keep the top-n highest similarities per document and then clustered using a combination of graph layout and average-link clustering. Cluster results from the nine similarity approaches were compared using (1) within-cluster textual coherence based on the Jensen-Shannon divergence, and (2) two concentration measures based on grant-to-article linkages indexed in MEDLINE.

Conclusions: PubMed's own related article approach (PMRA) generated the most coherent and most concentrated cluster solution of the nine text-based similarity approaches tested, followed closely by the BM25 approach using titles and abstracts. Approaches using only MeSH subject headings were not competitive with those based on titles and abstracts.

Figure 1. Cluster size distributions for the nine similarity approaches.

Figure 2. Textual coherence values by cluster size for the nine similarity approaches.

Coherence is a measure of cluster quality. A higher value of coherence indicates a higher degree of textual similarity between the titles and abstracts within a cluster than does a lower value of coherence. Data are shown for cluster size bins of at least 15 clusters.

Figure 3. Precision-recall curves for each cluster solution based on grant-to-article linkages.

To calculate precision-recall, clusters are first ordered by the fraction of articles referencing an NIH grant. Precision is the cumulative fraction of articles referencing the NIH grants, while recall is the cumulative fraction of articles in the cluster solution.

Figure 4. Two-dimensional map of the PMRA cluster solution, representing nearly 29,000 clusters and over two million articles.

The map was generated with cluster-to-cluster similarity values using the DrL graph layout routine . Color legend: Chemistry (blue), Engineering (cyan), Biology (green), Biotechnology (teal), Infectious Disease (brick red), Medical Specialties (red), Health Services (peach), Brain (orange), Social Sciences (yellow), Computer Sciences (pink).