Online reliable semi-supervised learning on evolving data streams

Abstract

In todays digital era, a massive amount of streaming data is automatically and continuously generated. To learn such data streams, many algorithms have been proposed during the last decade. Due to the dynamic nature of streaming data, the learning algorithms must be adaptive to handle concept drift and work under limited memory and time. Currently, most existing works assume that the true class labels of all incoming instances are immediately available. In real-world applications, labeling every data item in data streams is time and resource consuming. A more realistic situation is that only a few instances in data streams are labeled. Thereby, how to design a new efficient and effective learning algorithm that can handle concept drift, label scarcity, and work under limited resources is of significant importance. In this paper, we propose a new online semi-supervised learning algorithm by modeling concept drifts with a set of micro-clusters. These micro-clusters are dynamically maintained to capture the evolving concepts with error-based representative learning. In this way, local concept drifts are captured more quickly and finally support effective data stream learning. Extensive experiments on several data sets demonstrate that our learning model allows yielding high classification performance compared to many state-of-the-art algorithms.