北京邮电大学学报

  • EI核心期刊

北京邮电大学学报 ›› 2019, Vol. 42 ›› Issue (3): 127-132.doi: 10.13190/j.jbupt.2018-151

• 研究报告 • 上一篇    

一种快速的特征选择框架和方法

仇利克1, 刘竞2, 孙中卫3, 赵扬帆4   

  1. 1. 山东外贸职业学院 信息管理系, 青岛 266100;
    2. 青岛农业大学 理学与信息科学学院, 青岛 266100;
    3. 青岛理工大学 信息与控制工程学院, 青岛 266100;
    4. 山东青岛烟草有限公司 综合计划处, 青岛 266100
  • 收稿日期:2018-07-04 出版日期:2019-06-28 发布日期:2019-06-20
  • 作者简介:仇利克(1979-),女,讲师,E-mail:qllike@163.com.
  • 基金资助:
    国家重点研发计划项目(2016YFC1401907);国家自然科学基金项目(61827810)

A Fast Feature Selection Framework and Method

QIU Li-ke1, LIU Jing2, SUN Zhong-wei3, ZHAO Yang-fan4   

  1. 1. Information Management Department, Shandong Foreign Trade Vocational College, Qingdao 266100, China;
    2. Science and Information College, Qingdao Agricultural University, Qingdao 266100, China;
    3. School of Information and Control Engineering, Qingdao University of Technology, Qingdao 266100, China;
    4. Comprehensive Planning Office, Shandong Qingdao Tobacco Copany Limited, Qingdao 266100, China
  • Received:2018-07-04 Online:2019-06-28 Published:2019-06-20

摘要: 针对特征选择过程中准确率和计算效率不平衡问题,提出了一种快速特征选择框架(FFFS).基于该框架,使用最小冗余最大相关方法(MRMR)选择候选特征,借助序列前向选择方法(SFS)验证性能,并通过限定迭代次数提高计算性能.与MRMR、SFS和混合序列浮动前向选择算法(FDHSFFS)的对比实验结果表明,提出的快速特征选择算法MRMR-SFS能在预测准确率和计算效率之间取得较好的平衡.

关键词: 特征选择, filter, wrapper, hybrid, 性能预测, 相关系数

Abstract: Aiming at the imbalance between accuracy and computational efficiency in feature selection, a fast feature selection framework (FFFS) is proposed. Based on this framework, a fast feature selection algorithm, MRMR-SFS, is proposed. The minimum redundancy maximum relevance (MRMR) method is used to select the candidate features, and sequential forward selection (SFS) method is used to verify the performance of the candidate features as well. It improves the calculation efficiency by limiting the number of iterations. Comparison experiments with the MRMR, SFS and a filter-dominating hybrid sequential floating forward selection algorithms demonstrate that MRMR-SFS can balance the accuracy and computational efficiency well.

Key words: feature selection, filter, wrapper, hybrid, performance prediction, correlation coefficient

中图分类号: