Transfer Learning for Semi-supervised Classification of Non-stationary Data Streams

In the scenario of data stream classification, the occurrence of recurring concept drift and the scarcity of labeled data are very common, which make the semi-supervised classification of data streams quite challenging. To deal with these issues, a new classification algorithm for partially labeled streaming data with recurring concept drift is proposed. CAPLRD maintains a pool of concept-specific classifiers and utilizes historical classifiers to label unlabeled data, in which the unlabeled data are labeled by a weighted-majority vote strategy, and concept drifts are detected by automatically monitoring the threshold of classification accuracy on different data chunks. The experimental results illustrate that the transfer learning from historical concept-specific classifiers can improve labeling accuracy significantly, the detection of concept drifts and classification accuracy effectively.