Graph analysis algorithms such as PageRank and HITS have been successful in Web environments because they are able to extract important inter-document relationships from manually-created hyperlinks. We consider the application of these techniques to biomedical text retrieval. In the current PubMed ® search interface, a MEDLINE ® citation is connected to a number of related citations, which are in turn connected to other citations. Thus, a MEDLINE record represents a node in a vast content-similarity network. This article explores the hypothesis that these networks can be exploited for text retrieval, in the same manner as hyperlink graphs on the Web.
We conducted a number of reranking experiments using the TREC 2005 genomics track test collection in which scores extracted from PageRank and HITS analysis were combined with scores returned by an off-the-shelf retrieval engine. Experiments demonstrate that incorporating PageRank scores yields significant improvements in terms of standard ranked-retrieval metrics.