.png){kind=icon}
.png){kind=icon}
Return
Rational linear kernelized weighted fuzzy rough attribute selection with class separability
.png){kind=icon}
PRE
AI
DOI：10.1016/j.fss.2025.109600.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Attribute selection is widely used in data mining to reduce dimensionality and computational overhead, which improves the generalization and efficiency of machine learning models. However, it is a challenging task to select the most representative subset of attributes from data characterized by uncertainty, sparsity, and heterogeneity. Most existing fuzzy rough set-based attribute selection methods focus on improving specific characteristics without a comprehensive consideration, limiting their effectiveness. Moreover, these methods overlook the class distribution information in data, which leads to poor representation of the selected attribute subset. Motivated by these issues, a Rational Linear kernelized weighted fuzzy rough attribute selection method with class separability (RLWAS-CS) is proposed in this paper. A Rational Linear (RL) kernelized fuzzy similarity relation, derived from the Rational Quadratic kernel and mixed attribute distance measure, is first defined to accurately capture sample similarities in sparse and heterogeneous space. On this basis, an RL kernelized weighted fuzzy rough set model (RLWFRS) is proposed. In this model, the weight comprehensively reflects the membership between the sample and each class in the complete attribute space, which improves the discriminability of fuzzy approximation membership and enhances its ability to handle uncertainty. Additionally, an attribute evaluation function is designed based on the RLWFRS model. It integrates class separability and captures both the class distribution characteristics and the dynamic relation between attribute separability and redundancy. Finally, the RLWAS-CS algorithm is presented for the attribute selection. Experimental results show that RLWAS-CS outperforms baseline methods in effectiveness.
Journal
IF:
2.7
Papers:
7.6K
Citations:
1.5W
View more organizations