Factor Analytic Approach to Transitive Text Mining using Medline Descriptors

Abstract

Matrix decomposition methods were applied to examples of noninteractive literature sets sharing implicit relations. Document-by-term matrices were created from downloaded PubMed literature sets, the terms being the Medical Subject Headings (MeSH descriptors) assigned to the documents. The loadings of the factors derived from singular value or eigenvalue matrix decomposition were sorted according to absolute values and subsequently inspected for positions of terms relevant to the discovery of hidden connections. It was found that only a small number of factors had to be screened to find key terms in close neighbourhood, being separated by a small number of terms only. It is concluded that in literature-based discovery processes the decomposition methods combined with human inspection of the created factors may play an important role provided MeSH descriptors are analysed.