An Improved EM-Based Semi-supervised Learning Method

The iterative process in the standard EM-based semi-supervised learning includes two steps: firstly, use the classifier constructed in previous iteration to classify all unlabeled samples; then, train a new classifier based on the reconstructed training set, which is composed of labeled samples and all unlabeled samples. There is a problem in the process of reconstructing the training set: unlabeled samples should be divided into two parts by the classifier, reliable and misclassified, the misclassified part is considered as training samples. This affects the finally classification performance and convergence efficiency. This paper presents an improved method, which reconstructs the labeled samples set and then considers it as the training set in each iterative process. The process is implemented by adding the reliable part in unlabeled samples to labeled samples and forming a new labeled samples set. Experimental results in text classification show that the proposed method improves the classification performance and convergence efficiency.