An Active Learning Method for Mining Time-Changing Data Streams

Many applications generate continuous, time-changing data streams. Mining it for an adaptive classifier is of great interest and challenge. Many previous efforts impractically assume the labeled data is available and can be mined at anytime. In this paper, we propose an effective active learning method to mine time-changing data streams efficiently. It designs a way to monitoring the possible changes on the fly without need knowing the labeled data. Upon the suspected changes are indicated, it employs a light-weight uncertainty sampling algorithm to choose the most informative instances to label. With these representative labeled instances, it tests the significance of the suspected changes. If the changes indeed cause significant performance deterioration of the current classifier, it reconstructs the old model. Thus, our method can reliably detect significant changes, quickly adapt to concept-drift, and result effective models. Experimental results from real-world data confirm the advantages of our method.