Integration of textual content and link information for accurate clustering of science fields

Abstract

Hybrid clustering methods that exploit both text and citations might achieve better results than pure text-based or link-based methods. We experiment with Fisher's inverse chi-square method to derive a hybrid solution to map the bioinformatics field. A set of documents published in a list of core bioinformatics journals is extracted from the Web of Science (WoS) and extended with bibliometrically related records. Pairwise distances between documents are converted to p-values with respect to randomized datasets and the inverse chi-square method is used to combine the p-values from both information sources. This method can handle distances stemming from metrics with different distributional characteristics and avoids domination of any data source. For a correct application of Fisher's method we also introduce a slight, rank-preserving modification to the formula for bibliographic coupling. We evaluate clustering results by the mean Silhouette coefficient and also assess the performance in a classification setting for which we construct a 'ground truth' based on the Medical Subject Headings (MeSH), annotated by experts. To retrieve the MeSH terms, each WoS document is matched against Medline. Latent semantic analysis and a hierarchical clustering technique are applied to the MeSH-by-document matrix to determine document clusters, which are then post-processed by an iterative shrinking technique to only retain well-defined categories. We compare clustering and classification performances of sheer text and citation-based methods, of Fisher's method and of other data integration schemes. In general, text is more powerful than cited references, and dimensionality reduction by SVD further improves results; however, the best outcome is obtained by integration.