ARM-stream: active recovery of miscategorizations in clustering-based data stream classifiers

• Published in: Data mining and knowledge discovery Vol. 39; no. 5; p. 47
• Main Authors: Cavalcanti, Douglas Monteiro, Cerri, Ricardo, Faria, Elaine Ribeiro
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.09.2025; Springer Nature B.V
• Subjects: Algorithms; Artificial Intelligence; Chemistry and Earth Sciences; Clustering; Computer Science; Data Mining and Knowledge Discovery; Data transmission; Decoupling; Error recovery; Information Storage and Retrieval; Labels; Modules; Physics; Recovery; Source code; Statistics for Engineering; Concept evolution; Data stream; Concept drift; Active learning
• ISSN: 1384-5810; 1573-756X
• DOI: 10.1007/s10618-025-01124-4

The non-stationary feature-class interaction problem arises in dynamic data stream environments, where online classifiers avoid frequent label requests by relying on feature-class interaction assumptions to obtain class information from unlabeled data, even though these very assumptions may become invalid as this interaction changes unexpectedly. Clustering-based data stream classifiers mitigate this by leveraging active learning strategies. However, because these strategies are integral to the update process, enhancing reliability within the classifier is typically limited to increasing labeling resources, diminishing the classifier’s ability to learn from unlabeled data. In this paper we tackle this problem by decoupling the error-handling concerns from the classifier into a new framework called ARM-Stream, which works as a fail-safe layer that aims to intervene only when the classifier’s update process would fail, otherwise preserving the classifier’s original label resource usage patterns. ARM-Stream’s architecture, defined through abstract modules, ensures easy integration across different classifiers and easy customization and extension of its modules. To the best of our knowledge, ARM-Stream is the first error recovery framework for clustering-based data stream classifiers. As a study case, we test the framework over three classifiers: MINAS, CDSC-AL, and ECHO, chosen for their differing update process. An off-the-shelf source code library for the framework is provided.