| Résumé | Semi-supervised learning has attracted much attention over the past decade because it provides the advantage of combining unlabeled data with labeled data to improve the learning capability of models. Cotraining is a representative paradigm of semi-supervised learning methods. Typically, some co-training style algorithms, such as co-training and co-EM, learn two classifiers based on two views of the instance space. But they have to satisfy the assumptions that these two views are sufficient and conditionally independent given the class labels. Other co-training style algorithms, such as multiple-learner, use two different underlying classifiers based on only a single view of the instance space. However, they could not utilize the labeled data effectively, and suffer from the early convergence. After analyzing various co-training style algorithms, we have found that all of these algorithms have symmetrical framework structures that are related to their constraints. In this paper, we propose a novel unsymmetrical-style method, which we call the unsymmetrical co-training algorithm. The unsymmetrical co-training algorithm combines the advantages of other co-training style algorithms and overcomes their disadvantages. Within our unsymmetrical structure, we apply two unsymmetrical classifiers, namely, the self-training classifier and the EM classifier, and then train these two classifiers in an unsymmetrical way. The unsymmetrical co-training algorithm not only avoids the constraint of the conditional independence assumption, but also overcomes the flaws of the early convergence and the ineffective utilization of labeled data.We conduct experiments to compare the performances of these co-training style algorithms. From the experimental results, we can see that the unsymmetrical co-training algorithm outperforms other co-training algorithms. |
|---|
