A novel Graph-based Selection Wrapper for learning enhancement in a semi-supervised manner

In real-world applications, labeled data are often inadequate while unlabeled data are available in large quantities. Many semi-supervised learning (SSL) algorithms are proposed to take advantage of unlabeled data. In this paper we propose a novel wrapper method for semi-supervised learning called graph-based selection wrapper (GSW), which aims to improve an existing classifier in a semi-supervised manner. A new way of deciding the confidence of unlabeled data is introduced. GSW uses labeled data points and confident unlabeled data points together with their labels estimated by the target classifier to retrain a new classifier. Experimental results for transmembrane protein datasets and several benchmark databases show that our method is more effective than other related approaches.