Semi-supervised Ensemble Learning of Data Streams in the Presence of Concept Drift

Increasing access to very large and non-stationary datasets in many real problems has made the classical data mining algorithms impractical and made it necessary to design new online classification algorithms. Online learning of data streams has some important features, such as sequential access to the data, limitation on time and space complexity and the occurrence of concept drift. The infinite nature of data streams makes it hard to label all observed instances. It seems that using the semi-supervised approaches have much more compatibility with the problem. So in this paper we present a new semi-supervised ensemble learning algorithm for data streams. This algorithm uses the majority vote of learners for the labeling of unlabeled instances. The empirical study demonstrates that the proposed algorithm is comparable with the state-of-the-art semi-supervised online algorithms.