Semi-supervised text categorization with only a few positive and unlabeled documents

This paper studies a special case of semi-supervised text categorization. We want to build a text classifier with only a set P of labeled positive documents from one class (called positive class) and a set U of a large number of unlabeled documents from both positive class and other diverse classes (called negative class). This kind of semi-supervised text classification is called positive and unlabeled learning (PU-Learning). Although there are some effective methods for PU-Learning, they do not perform very well when the labeled positive documents are very few. In this paper, we propose a refined method to do the PU-Learning with the known technique combining Rocchio and K-means algorithm. Considering the set P may be very small (≤5%), not only we extract more reliable negative documents from U but also enlarge the size of P with extracting some most reliable positive documents from U. Our experimental results show that the refined method can perform better when the set P is very small.