A concept drift-oriented active learning method

Abstract

The continuous growth and evolving internal distribution of streaming data were long challenged by the phenomenon of concept drift in classification tasks. In practical applications, labeling costs were prohibitively high, and active learning was frequently employed to mitigate the scarcity of labels. However, in drifting environments, a single sampling strategy was prone to induce selection bias and information redundancy. In this paper, a concept drift-oriented active learning method, termed CDAL, was proposed, which integrated drift detection with a clustering-based representative sampling mechanism. Upon detecting a distribution change, a candidate set was constructed from neighboring samples and structurally partitioned using online clustering. A global average uncertainty was then calculated as a dynamic reference, and the overall uncertainty level within each cluster was compared. Based on this comparison, either a centroid sample or a random sample was adaptively selected from the cluster for labeling. This design eliminated the need for preset fixed thresholds and balanced the information content of samples with their distribution coverage while controlling annotation costs. Experiments conducted on multiple real-world and synthetic datasets demonstrated that CDAL achieved superior cumulative accuracy and related evaluation metrics compared to baseline methods and was capable of rapidly restoring classification performance for new concepts after drift occurred, thereby validating the effectiveness of the proposed strategy.