Semi-supervised learning of text classification on bacterial protein-protein interaction documents

Protein-protein interaction (PPI) network is essential to understand the fundamental processes governing cell biology. The mining and curation of PPI knowledge is critical for analyzing high-throughput genomics and proteomics data. Several PPI knowledge bases have been generated through expensive manual curation but far from comprehensive. It is desired to have a document classification system which can classify documents as PPI-related or not PPI-related and therefore assist the mining and curation of PPI knowledge. In order to build document classification systems, an annotated corpus is needed where each document in the corpus is tagged with a label (either positive or negative). However, it is usually the case that only a small number of positive documents can be obtained manually or from existing PPI knowledge bases with literature evidences. Meanwhile, there are a large number of unlabeled documents where most of them are not PPI-related. Machine learning based on a small number of positives and a large number of unlabeled documents is called learning from positive and unlabelled documents (LPU) which has been studied in the general domain. A popular approach for LPU is a two-step strategy where the first step is to obtain reliable negative documents (RN) and the second step is to refine RN using various methods such as clustering or boosting. In this paper, we tackle the problem of LPU for PPI document classification and compare three two-step procedures based on a public data set, Reuters-21578. One is to obtain a negative data set by building a machine learning classifier which treats each unlabelled document as negatives and then classifies unlabelled documents. The second procedure is to refine the negative data set iteratively and consider those unlabeled documents always classified as negative as reliable negatives. The third procedure is to augment the negative data set iteratively by including unlabeled documents classified as negative in any iteration. Three machine learning algorithms were deployed for each two-step procedure.