A Robust Ensemble Classification for Drifting Concepts

The concept drift phenomenon in a data stream mining refers to any circumstances under which the target concept changes over time abruptly or gradually. A machine learning algorithm should be equipped with some instruments in order to track such changing concepts whatever its extent or rate of drift is. In this article, we introduce a regression-based ensemble learning algorithm which is robust to outliers and fastly adapts to any concept drift as well. Fast adpatation to concept drift is materialized by regression-type combiners and robustness is facilitated into the algorithm by utilizing a 少一 likelihood estimation. Some simulation results with artificial data sets with concept drift verify that the proposed method shows a good classification accuracy under concept drift and various level of noises.