Sample Selection Method for Concept Drift

The concept drift problem makes it impossible to achieve stable predictions in a non-stationary environment. This is because the occurrence of concept drift causes the old samples to become noisy, resulting in the knowledge learned by the model no longer being applicable in the new environment. Existing methods have difficulty in identifying the region where the drift occurs, and thus cannot select the appropriate samples for learning. In this paper, we propose a dynamic sample selection method that first uses a micro cluster-based clustering algorithm to partition the feature space into regions, and then performs sample denoising in the region where the new samples are located, while reusing the removed samples with high weights. The denoising effect of the algorithm was verified on four synthetic datasets containing different drift types, and the validity of the results after sample reduction was tested on real datasets. The experimental results show that the method can effectively remove the drift noise caused by the concept drift, and can inductively summarize the more representative sample points for model learning, which improves the performance of the prediction model on streaming data.