Concept Drift Detection of Positive Class in Skewed Data Streams
Zhang Yuhon · Jisuanji kexue yu tansuo · 2013
The concept drift is common in skewed data stream(SDS). However, the most detection algorithms of concept drift assume that the class distributions of data streams are balanced, and are not suitable in skewed data streams. Therefore, this paper proposes a detection approach for concept drifts in SDS, called CDPSD. Firstly, it adopts the modified resample method, which makes the instances in different concepts belong to different data blocks, and then builds the classifiers. Secondly, it uses the class distribution of the positive not all instances to detect the concept drifts and modify the classifiers. The experiments show that CDPSD can detect the concept drift, update the classifier in time, and promote the classification results of positive instances.