支持向量机递归特征消除法选择变量

1.数据

rm(list = ls())
library(caret)
library(mlbench)
data(Sonar)
dat = Sonar
#dat$Class = ifelse(dat$Class=="M",0,1)
#dat$Class = factor(dat$Class)
dat[1:4,1:4]
##       V1     V2     V3     V4
## 1 0.0200 0.0371 0.0428 0.0207
## 2 0.0453 0.0523 0.0843 0.0689
## 3 0.0262 0.0582 0.1099 0.1083
## 4 0.0100 0.0171 0.0623 0.0205
dim(dat)
## [1] 208  61
table(dat$Class)
## 
##   M   R 
## 111  97

这个数据是mlbench里的内置数据。前207列是变量，最后一列是分类，因子型的。

2.拆分训练集和测试集

set.seed(123)
trainIndex <- createDataPartition(dat$Class, p = 0.7, list = FALSE)
train <- dat[trainIndex, ]
test <- dat[-trainIndex, ]

3.训练模型

需要的时间有点长，所以设置了只运行一次的跳过机制。

# Define the control parameters for SVM-RFE
ctrl <- rfeControl(functions = caretFuncs, 
                   method = "cv", number = 10)

# Train the SVM model using SVM-RFE
set.seed(123)
f = "svmsonar.Rdata"
if(!file.exists(f)){
  svm_rfe_model <- rfe(x = train[, -ncol(train)], 
                     y = train$Class, 
                     rfeControl = ctrl, method = "svmRadial")
  save(svm_rfe_model,file = f)
}
load(f)

4.查看模型选择的变量，计算C-index

看模型选中的变量有哪些

svm_rfe_model$optVariables
##  [1] "V11" "V12" "V13" "V10" "V9"  "V49" "V51" "V36" "V48" "V35" "V37" "V52"
## [13] "V47" "V5"  "V4"  "V21" "V46" "V43" "V14" "V8"  "V1"  "V20" "V22" "V6" 
## [25] "V34" "V23" "V44" "V45" "V50" "V24" "V3"  "V31" "V38" "V57" "V58" "V19"
## [37] "V25" "V53" "V42" "V28" "V33" "V27" "V26" "V2"  "V32" "V59" "V7"  "V54"
## [49] "V15" "V55" "V56" "V29" "V16" "V60" "V40" "V17" "V39" "V18" "V30" "V41"

计算C-index值

# Evaluate the performance of the model using ROC curve
library(pROC)
svm_rfe_pred <- predict(svm_rfe_model, newdata = test[, -ncol(test)])

library(Hmisc)
rcorr.cens(as.numeric(test$Class),svm_rfe_pred)
##        C Index            Dxy           S.D.              n        missing 
##   8.239583e-01   6.479167e-01   9.587715e-02   6.200000e+01   0.000000e+00 
##     uncensored Relevant Pairs     Concordant      Uncertain 
##   6.200000e+01   1.920000e+03   1.582000e+03   0.000000e+00