MinoRare: Handling Rare Minority Instances in Imbalanced and Drifting Data Streams Through Adaptive Instance Weighting Scheme
Hina Farooq, Muhammad Usman, Huanhuan Chen · IEEE Transactions on Big Data · 2025
The joint occurrence of concept drifts and class imbalance poses a significant challenge to the classification tasks in data streams. Among the various challenges in imbalanced data, rare minority instances deeply embedded within the majority class space present a unique difficulty. These instances are not only few in number but are also surrounded by majority class data, making the joint problem even more complex to address. This paper proposes MinoRare, a batch-based ensemble technique, to address the joint problem with a special focus on rare minority instances. MinoRare divides the batch data into subframes and applies weight instance criteria at the subframe level. Weights are assigned to minority and majority class instances based on their difficulty level, which is calculated by analyzing the neighborhood density of each instance in a subframe coupled with the global class overlap calculated at batch-level. Additionally, the ensemble pool is kept updated with classifiers setup with new weighted instance data, whereas outdated classifiers are removed making it adaptive to recent concepts and the current state of imbalanced data. By mitigating the rarity issue, MinoRare ensures that classifiers focus on critical rare minority instances through a targeted instance-weighting strategy, enhancing their ability to learn and represent decision boundaries for the minority class effectively, and ultimately improving classification performance in evolving data streams. Experiments against 13 state-of-the-art methods on 345 data streams containing a variety of concept drift and class imbalance coupled with different data difficulty factors demonstrate the efficacy of the proposed method.