R語(yǔ)言利用caret包比較ROC曲線

我們之前探討了多種算法，每種算法都有優(yōu)缺點(diǎn)，因而當(dāng)我們針對(duì)具體問(wèn)題去判斷選擇那種算法時(shí)，必須對(duì)不同的預(yù)測(cè)模型進(jìn)行重做評(píng)估。為了簡(jiǎn)化這個(gè)過(guò)程，我們使用caret包來(lái)生成并比較不同的模型與性能。
操作
加載對(duì)應(yīng)的包與將訓(xùn)練控制算法設(shè)置為10折交叉驗(yàn)證，重復(fù)次數(shù)為3：
library(ROCR)
library(e1071)
library("pROC")
library(caret)
library("pROC")
control = trainControl(method = "repaetedcv",
                       number = 10,
                       repeats =3,
                       classProbs = TRUE,
                       summaryFunction = twoClassSummary)

使用glm在訓(xùn)練數(shù)據(jù)集上訓(xùn)練一個(gè)分類器

glm.model = train(churn ~ .,
                  data= trainset,
                  method = "glm",
                  metric = "ROC",
                  trControl = control)
使用svm在訓(xùn)練數(shù)據(jù)集上訓(xùn)練一個(gè)分類器

svm.model = train(churn ~ .,
                  data= trainset,
                  method = "svmRadial",
                  metric = "ROC",
                  trControl = control)
使用rpart函數(shù)查看rpart在訓(xùn)練數(shù)據(jù)集上的運(yùn)行情況

rpart.model = train(churn ~ .,
                    data = trainset,
                    method = "svmRadial",
                    metric = "ROC",
                    trControl = control)
使用不同的已經(jīng)訓(xùn)練好的數(shù)據(jù)分類預(yù)測(cè)：

glm.probs = predict(glm.model,testset[,!names(testset) %in% c("churn")],type = "prob")
svm.probs = predict(svm.model,testset[,!names(testset) %in% c("churn")],type = "prob")
rpart.probs = predict(rpart.model,testset[,!names(testset) %in% c("churn")],type = "prob")
生成每個(gè)模型的ROC曲線，將它們繪制在一個(gè)圖中：

glm.ROC = roc(response = testset[,c("churn")],
              predictor = glm.probs$yes,
              levels = levels(testset[,c("churn")]))
plot(glm.ROC,type = "S",col = "red")

svm.ROC = roc(response = testset[,c("churn")],
              predictor = svm.probs$yes,
              levels = levels(testset[,c("churn")]))
plot(svm.ROC,add = TRUE,col = "green")

rpart.ROC = roc(response = testset[,c("churn")],
              predictor = rpart.probs$yes,
              levels = levels(testset[,c("churn")]))

plot(rpart.ROC,add = TRUE,col = "blue")

三種分類器的ROC曲線

說(shuō)明

將不同的分類模型的ROC曲線繪制在同一個(gè)圖中進(jìn)行比較，設(shè)置訓(xùn)練過(guò)程的控制參數(shù)為重復(fù)三次的10折交叉驗(yàn)證，模型性能的評(píng)估參數(shù)為twoClassSummary,然后在使用glm,svm,rpart,三種不同的方法建立分類模型。
從圖中可以看出，svm對(duì)訓(xùn)練集的預(yù)測(cè)結(jié)果（未調(diào)優(yōu)）是三種分類算法里最好的。