Concept Drift Adaptive Prediction Method Based on Dynamic Sample Selection

Concept drift is an important performance factor in stream data mining, mainly handled by incremental updating or retraining models, but not fully utilizing existing knowledge. This paper proposed an concept drift adaptive prediction method based on dynamic sample selection, starting from the comprehensive use of all samples. The method performs local consistency based drift detection when new samples arrive, removes noisy samples in the region when drift is detected, and reuses historically similar concepts when new concepts are detected. Finally, multi-representative point summarization is performed for different categories of samples in the region, and the prediction model is updated simultaneously. In this paper, the denoising effect is verified on synthetic datasets containing different drift types, and the prediction task is performed on the real dataset. The experimental results show that the method can effectively remove the drift noise due to conceptual drift, which effectively improves the performance of the prediction model. The prediction outperforms the popular concept drift adaptive model. © 2024 Chinese Institute of Electronics. All rights reserved.