caret包应用之四：模型预测与检验

模型建立好后，我们可以利用predict函数进行预测，例如预测检测样本的前五个

predict(gbmFit1, newdata = testx)[1:5]

为了比较不同的模型，还可用装袋决策树建立第二个模型，命名为gbmFit2

gbmFit2= train(trainx, trainy,method = "treebag",trControl = fitControl)
models = list(gbmFit1, gbmFit2)

另一种得到预测结果的方法是使用extractPrediction函数，得到的部分结果如下显示

predValues = extractPrediction(models,testX = testx, testY = testy)
head(predValues)

     obs     pred model dataType  object
1 Active   Active   gbm Training Object1
2 Active   Active   gbm Training Object1
3 Active Inactive   gbm Training Object1
4 Active   Active   gbm Training Object1
5 Active   Active   gbm Training Object1 

6 Active Active gbm Training Object1

从中可提取检验样本的预测结果

testValues = subset(predValues, dataType == "Test")

如果要得到预测概率，则使用extractProb函数

probValues = extractProb(models,testX = testx, testY = testy)
testProbs = subset(probValues, dataType == "Test")

对于分类问题的效能检验，最重要的是观察预测结果的混淆矩阵

Pred1 = subset(testValues, model == "gbm")
Pred2 = subset(testValues, model == "treebag")
confusionMatrix(Pred1$pred, Pred1$obs)
confusionMatrix(Pred2$pred, Pred2$obs)

结果如下，可见第一个模型在准确率要比第二个模型略好一些

          Reference
Prediction Active Inactive
  Active       65       12
  Inactive      9       45
                                      
               Accuracy : 0.8397          

Reference

Prediction Active Inactive
  Active       63       12
  Inactive     11       45
                                          
               Accuracy : 0.8244          

最后是利用ROCR包来绘制ROC图

prob1 = subset(testProbs, model == "gbm")
prob2 = subset(testProbs, model == "treebag")
library(ROCR)
prob1$lable=ifelse(prob1$obs=='Active',yes=1,0)
pred1 = prediction(prob1$Active,prob1$lable)
perf1 = performance(pred1, measure="tpr", x.measure="fpr" )
plot( perf1 )