期刊
ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE
卷 115, 期 -, 页码 -出版社
PERGAMON-ELSEVIER SCIENCE LTD
DOI: 10.1016/j.engappai.2022.105226
关键词
Fuzzy rough sets; Feature selection; Divergence measure; Dependency degree and feature significance; Three-level analysis; Double quantification
类别
资金
- National Natural Science Foundation of China [61673285, 11671284]
- Sichuan Science and Technology Program of China [2021YJ0085, 2022NSFSC0929]
- Laurent Mathematics Center of Sichuan Normal University [ZD20220101]
- National-Local Joint Engineering Laboratory of System Credibility Automatic Verification [ZD20220101]
Feature selection plays a crucial role in machine learning and knowledge acquisition. This study proposes a divergence-based fuzzy rough sets method and introduces double-quantitative feature selection algorithms to improve the existing algorithms and achieve better classification performance.
Feature selection benefits machine learning and knowledge acquisition, and it usually resorts to various intelligent methodologies. Fuzzy rough sets act as a powerful platform of intelligent processing, and they have introduced divergence measures to generate an effective method of feature selection, called FS-DD. However, Algorithm FS-DD still has advancement space, because its underlying dependency degree with absoluteness lacks decision-categorical manifestations and exhibits loose informatization. Within the framework of divergence-based fuzzy rough sets (Div-FRSs), we implement bidirectional three-level dependency measurements to establish double-quantitative feature selection, and two novel approaches of feature selection (i.e., Algorithms FS-AFS and FS-RFS) are designed to reconstruct and improve current Algorithm FS-DD. Based on divergence and lower-approximation matrices, we first make three-level measurements in vertical and horizontal directions, and correspondingly generate absolute and relative dependency degrees. Then, double-quantitative dependency degrees naturally induce double-quantitative feature significances, and the two types of uncertainty measures respectively exhibit granulation monotonicity and non-monotonicity. Furthermore, double-quantitative feature significances are utilized to motivate double-quantitative selection algorithms, i.e., absolute FS-AFS and relative FS-RFS. Finally, measurement properties and selection algorithms are fully validated by table examples and data experiments. This study systematically reveals hierarchical constructions and quantitative characteristics of dependency measurements in Div-FRSs, and the relative measures effectively extract class-specific and condensed information. For related selection algorithms, FS-AFS interprets existing FS-DD, while new FS-RFS outperforms the two to acquire better classification performances, as experimentally verified.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据