MinoClust: Exploiting Minority Sub-Cluster Dynamics for Classification in Imbalanced and Drifting Data Streams

The coexistence of concept drift and class imbalance in data streams poses significant challenges for classification tasks, particularly when compounded by additional data difficulty factors such as local drifts, high imbalance ratio, rare minority instances, and overlapping class boundaries. Conventional learning techniques ignore local drifts, i.e. within-class subcluster transitions leading to suboptimal performance. This paper introduces MinoClust, a batch-based classification ensemble method specifically designed to address the dynamic nature of sub-cluster transitions in the minority class. MinoClust clusters incoming data at the batch level and identifies trending dense regions in the minority space for targeted resampling. The ensemble is dynamically updated by incorporating classifiers trained on resampled data from these regions while removing underperforming classifiers. Experimental evaluations on 37 datasets, encompassing diverse concept drift scenarios and class imbalance settings with various data difficulty factors, demonstrate that explicitly addressing local drifts enhances classification performance in imbalanced and drifting data streams.