ADAPTIVE DATA REUSE FOR CLASSIFYING IMBALANCED AND CONCEPT-DRIFTING DATA STREAMS

Mining data streams has recently been the subject of extensive research efforts. However, most of the works conducted in this field assume a balanced class distribution underlying data streams. In this paper, therefore, we propose a new method for learning from imbalanced data streams. To deal with the problem of class imbalance, we select and reuse past data to improve the representation of the minority class. Different from previous methods, our method has the ability to automatically adapt data selection for concept drift. A data stream may experience a complicated concept drift, making data selection more difficult. Therefore, we consider several different candidate solutions of data selection, each of which is possibly more appropriate for certain data streaming conditions. In other words, no one of them is the best at all times. We make comparisons and identify the best candidate solution by cross-validation on the most recent training data. By experimental evaluations on simulated and real-world data streams, we show that our method achieves better performance than previous methods, especially when concept drift occurs.