Clustering-training for Data Stream Mining

Mining data streams has attracted much attention recently. Labeled samples needed by most current stream classification methods are more difficult and expensive to obtain than unlabeled ones. This paper proposed a semi-supervised learning algorithm - clustering-training to utilize the unlabeled samples. It uses clustering to select confidently unlabeled samples, and uses them to re-train the classifier incrementally. Experiments on synthetic and real data set showed the effectiveness of the proposed algorithm