【作者】 周密；

【导师】 张传林；

【作者基本信息】 暨南大学 ， 应用数学， 2016， 硕士

【摘要】 特征选择是机器学习领域研究的热点之一。本文通过介绍特征选择的相关背景及研究意义,分析一些特征选择方法的优缺点,利用过滤式和封装式这两类特征选择算法的互补性,提出一种基于Fisher Score及遗传算法的混合式特征选择方法。该方法先对所有特征的Fisher Score作一个线性变换,再利用变换后的Fisher Score生成遗传算法的初始种群,接着借鉴精英保留策略,用遗传算法的后续运算选出特征。以Sonar,WDBC,Arrhythmia,Hepatitis这四个数据集作为实验数据,用该方法分别选出它们的特征子集,再依据所选特征子集对原数据集降维,用1-最近邻分类器对降维后的样本分类,通过10重交叉验证法分别获得72.36%,95.64%,72.04%和87.83%的分类准确率,并且所需的迭代次数较少,特征选择的综合效果基本优于Fisher Score法(FS)、遗传算法(GA)和Fisher Score+遗传算法(FSGA)这3种对比方法,同时该方法能很好地剔除冗余特征,选出具有较高鉴别力的特征,是一种有效的特征选择方法。更多还原

【Abstract】 Feature selection is one of the popular researches in machine learning area. This article provides a brief review about the background of feature selection, and analyses the advantages and drawbacks of some feature selection methods. Due to the complementary of filter based and wrapper based feature selection algorithm, we propose a hybrid feature selection method. In the first place, the Fisher scores of all features will be mapped into a specific interval by a linear function, and then the rescaled Fisher scores will be utilized to generate the initial population of genetic algorithm. Finally, the initial population will be used in the subsequent procedure of genetic algorithm to perform feature selection with elitist strategy for reference. In this paper, we choose four data sets of Sonar,WDBC, Arrhythmia and Hepatitis to test the performance of our proposed algorithm. Feature subsets of the four data sets will be selected by our algorithm, and then the dimensionality of data sets will be reduced according to the selected feature subsets respectively. 1-NN classifier is used to classify the dimensionality reduced data sets, and respectively achieving the classification accuracy of 72.36%, 95.64%, 72.04% and 87.83% with ten-fold cross validation method. The experiment results show that, compared to the performance of Fisher Score(FS), Genetic Algorithm(GA) and Fisher Score Genetic Algorithm(FSGA), our algorithm is fit for eliminate redundant features, and it can select discriminative features. Above all, our method is effective in feature selection.更多还原