The semi-supervised learning techniques use abundant unlabeled data for helping to learn a better classifier if the number of instances is very less. A basic technique is to choose and label the unlabeled instances that the present classifier has higher classification confidence for enlarging the labeled training set and the updating the classifier, which is mostly utilized in two different paradigms of semi-supervised learning namely: co-training and self-training. But the actual labeled instances will be more reliable compared to self-labeled instances which would be labeled by a classifier. If unlabeled instances are assigned to wrong labels then the classification accuracy of classifier might be jeopardized. In this paper, a new instance selection technique is presented based on real labeled data. This will consider present classifier performance on unlabeled data as well as its performance only on real labeled data. In every iteration, this utilizes the accuracy changes in newly learned classifier over original labeled data as a criteria for deciding either the chosen most confident unlabeled instances would be accepted by further iteration or not. The experiments will be conducted in co-training as well as self-training while using Naïve Bayes (NB) as a base classifier. The results show that, SIS will significantly improve accuracy and classification of self-training and co-training. From results it can observe that it will improve the accuracy, Precision, Recall and F1 score compared with semi-supervised classification method.

Supervised Instance Selection (SIS), Data mining, Meta-learning, Algorithm selection.