<p>Concept drift is one of the main challenges in online learning, and its difficulty increases when the data stream is highly imbalanced—especially when drifts appear in minority class samples. Existing ensemble methods provide partial solutions but often rely on storing previous data chunks, incur high computational costs, or adapt slowly to evolving concepts.</p><p>In this work, we address the problem of real-time binary classification in imbalanced, non-stationary data streams and propose EFR-IC, an Ensemble Fuzzy Association Rule-based classifier designed to perform accurate, scalable learning capable of adapting to evolving concepts while giving additional attention to minority class patterns without storing any historical data, unlike existing ensemble methods that require historical chunks for adaptation.</p><p>The model operates in four stages. First, fuzzy association rules are extracted from the current chunk. Second, drift is detected using outlier-based rule-matching and cluster-variation criteria. Third, ensemble members are updated or replaced based on recent data, and finally, predictions are made via weighted voting. The integrated pipeline supports parallel processing for high-performance computing environments while enabling rapid adaptation to sudden, gradual, incremental, and recurrent drifts.</p><p>EFR-IC maintains stability under both stationary and drifting conditions through synchronized, chunk-based processing and adaptive ensemble management. Experimental results on synthetic and real-world datasets demonstrate competitive and often superior performance, achieving slight improvements in G-mean across several datasets, while remaining comparable to state-of-the-art methods overall. Moreover, the computationally intensive stages of rule extraction, genetic tuning, and ensemble updating are inherently parallelizable on multi-core CPUs or distributed clusters, enabling scalable deployment on high-performance computing platforms for real-time stream processing.</p>