An Evaluation of Two-Step Techniques for Positive-Unlabeled Learning in Text Classification

Abstract

Positive-unlabeled (PU) learning is a learning problem which uses a semi-supervised method for learning. In PU learning problem, the aim is to build an accurate binary classifier without the need to collect negative examples for training. Two-step approach is a solution for PU learning problem that consists of tow steps: (1) Identifying a set of reliable negative documents. (2) Building a classifier iteratively. In this paper we evaluate five combinations of techniques for two-step strategy. We found that using Rocchio method in step 1 and Expectation-Maximization method in step 2 is most effective combination in our experiments.