Instance Selection in Semi-supervised Learning

Abstract

Semi-supervised learning methods utilize abundant unlabeled data to help to learn a better classifier when the number of labeled instances is very small. A common method is to select and label unlabeled instances that the current classifier has high classification confidence to enlarge the labeled training set and then to update the classifier, which is widely used in two paradigms of semi-supervised learning: self-training and co-training. However, the original labeled instances are more reliable than the self-labeled instances that are labeled by the classifier. If unlabeled instances are assigned wrong labels and then used to update the classifier, classification accuracy will be jeopardized. In this paper, we present a new instance selection method based on the original labeled data (ISBOLD). ISBOLD considers not only the prediction confidence of the current classifier on unlabeled data but also its performance on the original labeled data only. In each iteration, ISBOLD uses the change of accuracy of the newly learned classifier on the original labeled data as a criterion to decide whether the selected most confident unlabeled instances will be accepted to the next iteration or not. We conducted experiments in self-training and co-training scenarios when using Naive Bayes as the base classifier. Experimental results on 26 UCI datasets show that, ISBOLD can significantly improve accuracy and AUC of self-training and co-training.