Online Active Learning for Evolving Data Streams with Limited Labels

The classification of data streams presents considerable challenges, especially in the context of concept drift and label scarcity. Many existing approaches struggle to efficiently classify data streams in the presence of concept drift, particularly when labeled data is scarce. This study proposes an innovative online active learning algorithm for data streams, referred to as OALEDS. The algorithm combines dynamic micro-clustering for concept representation with local class entropy modeling to effectively identify essential concepts and prototypes within continuously evolving data streams. Specifically, the algorithm places greater emphasis on prototypes in regions susceptible to concept drift, thereby concentrating on areas that reflect current data patterns. Additionally, a micro-clustering-based local class entropy strategy is employed to identify the most informative instances for labeling, optimizing the labeling process. Comprehensive experiments performed on both real-world and synthetic imbalanced datasets reveal that OALEDS efficiently adapts to concept drift, dynamically captures significant prototypes, and consistently surpasses state-of-the-art algorithms, even in scenarios with restricted labeling availability.