Stable region enhanced online learning method for intermediate verification latency and concept drift

In non-stationary data streams, the challenges of concept drift are compounded by Intermediate Verification Latency (IVL), which refers to the delay between the arrival of data features and their corresponding labels. This paper addresses IVL from a novel perspective by examining stable regions within the feature space, where data remain unaffected by concept drift. We propose Centroid-based Drift Index (CDI), an unsupervised metric that quantifies drift to identify these stable regions. Building upon this, we introduce Data Utilization informed by STable regions (DUST), a framework that effectively utilizes both temporarily unlabeled and delayed labeled data by distinguishing stable region data through micro-clusters. Comprehensive experiments conducted on synthetic and real-world datasets validate the effectiveness of CDI and DUST. The results demonstrate that DUST yields accuracy improvements ranging from 0.59% to 2.29% across various base models, with an average accuracy of 84.10% on synthetic streams and 63.18% on real-world streams. The source code and supplementary material are available at https://github.com/ZeroZill/CDI_DUST.