An Efficient Active Bayes Classifier Using Affinity Propagation on Unlabeled Data

When the labeled data are few, exploiting amount of unlabeled data can be helpful for improve learning performance of classifier. The key issue for active learning to solve is how to select the most ¿valuable¿ training samples to reduce labeled cost of amount of unlabeled samples. In the paper, we propose an efficient active Bayes classifier by using affinity propagation (AP) to select the most valuable samples on the unlabeled data sets. Firstly, the method clusters amount of unlabeled sample data by using AP algorithm. Secondly, through exemplars obtained from the results of cluster and similarity matrix, the method selects the unlabeled samples for experts to label. The method choose two kinds of most valuable instances to label: one is cluster center instance, which also called exemplar instance. Another is the instance which stays in the farthest position to the cluster center in the same cluster, that is, the sample which sets in the nearest position to the cluster boundary. The experiment results on UCI data sets demonstrate that our method is simple and more efficient compared with traditional active learning method.