New Drift Detection Method for Data Streams

摘要

Correctly detecting the position where a concept begins to drift is important in mining data streams. In this paper, we propose a new method for detecting concept drift. The proposed method, which can detect different types of drift, is based on processing data chunk by chunk and measuring differences between two consecutive batches, as drift indicator. In order to evaluate the proposed method we measure its performance on a set of artificial datasets with different levels of severity and speed of drift. The experimental results show that the proposed method is capable to detect drifts and can approximately find concept drift locations.