Semi-Supervised Self-Training Classification Based on Neighborhood Construction

Using the unlabeled data in the semi-supervised learning can significantly improve the accuracy of supervised classification. But in some cases, it may dramatically reduce the accuracy of the classification. The reason of such degradation is incorrect labeling of unlabeled data. In this article, we propose the method for high confidence labeling of unlabeled data. The base classifier in the proposed algorithm is the support vector machine. In this method, the labeling is performed only on the set of the unlabeled data that is closer to the decision boundary from the threshold. This data is called informative data. the adding informative data to the training set has a great effect to achieve the optimal decision boundary if the predicted label is correctly. The Epsilon- neighborhood Algorithm (DBSCAN) is used to discover the labeling structure in the data space. The comparative experiments on the UCI dataset show that the proposed method outperforms than some of the previous work to achieve greater accuracy of the self-training semi-supervised classification.