A new dictionary-based positive and unlabeled learning method

摘要

Positive and unlabeled learning (PU learning) is designed to solve the problem that we only utilize the labeled positive examples and the unlabeled examples to train a classifier. A variety of methods have been proposed to solve this problem by incorporating unlabeled examples into learning. However, many methods treat the original features as input in the training stage and then build the classifier. In this paper, by use of two-step strategy, a novel method with dictionary learning is proposed for PU learning, which is briefly called PUDL. The proposed method is done in two steps. Firstly, we extract reliable negative examples from unlabeled examples to form the negative class. We then utilize dictionary learning to construct the feature representation which can learn new features for the original input. Secondly, we propose RankSVM-based (Ranking Support Vector Machine) model to incorporate the positive class, extracted negative class and similarity weights into learning so as to improve performance of the classifier. The Lagrange multiplier method is applied to convert the original model into its dual form. In addition, we put forward an interactive optimization framework to optimize the proposed objective model and obtain the classifier. Finally, we verify the performance of PUDL via experiment and the results demostrate that PUDL performs better than state-of-the-art PU learning methods.