.png){kind=icon}
.png){kind=icon}
Return
Multi-objective sparrow search-based feature selection via weighted oversampling with intuitionistic fuzzy C-means for imbalanced classification
.png){kind=icon}
PRE
AI
L
S
W
J
DOI：10.1016/j.swevo.2026.102482.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
For the task of classifying imbalanced data, the poor recognition of minority (positive) samples leads to a significant degradation in classifier efficiency. Existing sampling methods are often developed without accounting for noise, and frequently fail to consider the underlying data distribution. Furthermore, multi-objective optimization strategies for feature selection often suffer from duplicate solutions in both the search and objective spaces. To overcome these challenges, we present a novel weighted oversampling scheme via intuitionistic fuzzy C-means clustering with symmetric relative entropy, along with a multi-objective Sparrow Search for feature selection incorporating intuitionistic fuzzy entropy. In first stage, employing the proposed symmetric relative entropy, we redefine the membership function, cluster center, and objective function of the intuitionistic fuzzy C-means scheme to improve its clustering efficacy. This improved clustering strategy is then applied to group minority and majority class samples in imbalanced dataset. Subsequently, we design a weighted oversampling method that takes into account inter-class distance, intra-class distance, and class size. Furthermore, we formulate the feature selection task as a multi-objective optimization problem by incorporating intuitionistic fuzzy entropy as the third objective. Finally, to enhance population diversity and achieve a better classification efficiency, we devise two environmental selection strategies to eliminate duplicate solutions across both the search space and the objective space for feature selection. Through optimization analyses using the CEC 2020 benchmark and comparative evaluations of various sampling and feature selection methods across 19 imbalanced datasets, our experiments confirm the superiority of the constructed methodology over other state-of-the-art methods across several metrics.
Journal
IF:
8.5
Papers:
2.1K
Citations:
1.0W
Organization
Cited Papers