Page 83 - 《软件学报》2020年第12期

张鑫 等:自然进化策略的特征选择算法研究 3749

Table 7 Classification accuracy and dimension reduction of MCC-NESand compared methods (3)
表 7 MCC-NES 及其对比算法的分类准确率和维度缩减率(3)
Spambase CA (%) DR (%) Classifier
MCC-NES 92.47(10-fold) 57.89 RBF-SVM
ACOFSS+mRMR 91.40(10-fold) 59.32 RBF-SVM
MCC-NES 93.84(10-fold) 49.12 CART
ACOFSS+mRMR 93.60(10-fold) 40.35 CART
Sonar CA (%) DR (%) Classifier
MCC-NES 96.83(70%~30%) 81.67 5-NN
FSFOA 86.98(70%~30%) 44.26 5-NN
PSO(4-2) 78.16(70%~30%) 81.26 5-NN
DE-FLS 76.77(70%~30%) 65.0 5-NN
MCC-NES 97.93(70%~30%) 78.0 1-NN
FSFOA 85.43(70%~30%) 57.37 1-NN
SVM-FuzCoc 73.17(70%~30%) 68.33 1-NN
SFS 66.43(50%-50%) 61.33 1-NN
SBS 62.20(50%-50%) 45.33 1-NN
SFFS 64.55(50%-50%) 61.33 1-NN
IFS-CoCo 85.70(10-flod) 57.17 1-NN
MCC-NES 81.20(2-fold) 90.0 RBF-SVM
FSFOA 65.86(2-fold) 54.09 RBF-SVM
HGAFS 87.02(2-fold) 75.0 RBF-SVM
ISEDBFO 92.80(10-flod) 56.66 RBF-SVM
ACBFO 93.50(10-flod) 56.5 RBF-SVM
MCC-NES 88.04(10-fold) 71.67 CART
FSFOA 82.69(10-fold) 52.45 CART
FS-NEIR 75.97(10-fold) 91.66 CART
Musk2 CA (%) DR (%) Classifier
MCC-NES 97.01(70%~30%) 80.28 1-NN
MCC-NES 96.92(70%~30%) 82.65 5-NN
DE-FLS 86.86(70%~30%) 78.67 5-NN
LSVT CA (%) DR (%) Classifier
MCC-NES 97.11(70%~30%) 97.0 1-NN
FSFOA 89.47(70%~30%) 98.71 1-NN
Rc-BBFA 94.60(70%~30%) 56.45 1-NN
SRBCT CA (%) DR (%) Classifier
MCC-NES 99.87(70%~30%) 98.62 1-NN
FSFOA 94.73(70%~30%) 49.06 1-NN
SVM-FuzCoc 98.88(70%~30%) 98.57 1-NN
Arcene CA (%) DR (%) Classifier
MCC-NES 97.50(70%~30%) 98.44 1-NN
FSFOA 88.33(70%~30%) 61.9 1-NN
Rc-BBFA 92.50(70%~30%) 48.66 1-NN
MCC-NES 98.33(70%~30%) 98.27 CART
FSFOA 73.69(70%~30%) 77.67 CART
UFSACO 67.40(70%~30%) 99.8 CART
RNA-Seq CA (%) DR (%) Classifier
MCC-NES 99.87(70%~30%) 99.46 1-NN
Rc-BBFA 94.53(70%~30%) 58.66 1-NN
Dorothea CA (%) DR (%) Classifier
MCC-NES 95.42(70%~30%) 99.58 1-NN
MCC-NES 96.53(70%~30%) 99.63 CART
FSFOA 95.83(70%~30%) 26.03 CART

在处理更高维的数据时,很多特征选择方法就显得无能为力了,比如在面对 LSVT,SRBCT,Arcene,RNA-Seq,
Dorothea 数据集时.我们在设定的最大迭代时间内(24 hours)内测试了本文提出的方法,并与最新提出的几个特
征选择方法 FSFOA,UFSACO,Rc-BBFA(该算法最适应于 1-NN 分类器,所以只选取了 1-NN 分类器的实验结果)
进行比较.实验结果表明:在处理高维问题时,我们算法的准确率几乎都是最优的,并且特征数量即使达到
100 000 时仍能给出令人满意的解.其中,在 Arcene 的 CART 分类器上更是比次优的准确率高出 25%,比
UFSACO 算法高出 30%左右.而且与 Rc-BBFA,FSFOA 相比,我们的算法在解决高维问题时的分类准确率和维
度缩减表现更为出色.