.png){kind=icon}
.png){kind=icon}
Return
SWARM: A novel Sample Weighting Approach for Rare Minority data classification in imbalanced and evolving data streams
.png){kind=icon}
PRE
AI
M
M
DOI：10.1016/j.knosys.2026.116781.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The joint occurrence of class imbalance and concept drift in streaming data creates a challenging learning problem for classification algorithms. Rare minority instances, which are scattered in the deep majority spaces and reside far away from dense minority areas, make learning more difficult in such scenarios. This paper proposes SWARM, a Sample Weighting Approach for Rare Minority data classification in imbalanced and drifting streams. The proposed method presents a novel instance weighting and resampling technique, WIRE (Weighted Instance Resampling Engine), that leverages misclassification cost for hard-to-learn minority instances, uses neighborhood analysis to guide a weighting policy that emphasizes rare minority data, and integrates a decay factor to make weights adaptable to the dynamic nature of data streams. It also employs a resampling strategy considering intra-class rarity and majority data to generate balanced training data batches. The ensemble pool is maintained and pruned under the R3D Ensemble component that adaptively manages the classifiers and prunes the pool through a multi-objective optimization method based on classifier diversities and weighted-recall performance. The experiments are performed on 57 data streams characterized by class imbalance, rare minority instances, and concept drifts. The comparative analysis against 10 state-of-the-art methods demonstrates the superiority of the proposed method, particularly in boosting recall performance and achieving higher G-Mean scores.
Keywords:
Data streams
Class imbalance
Rare minority data
Concept drifts
Ensemble learning
Journal
K
IF:
7.6
Papers:
1.2W
Citations:
4.5W
Organization
No organization information available
Cited Papers