在 R（插入符号）中重新运行 preProcess()、predict() 和 train() 时模型精度不同

Question

下面的数据只是一个例子，它是对这个或任何我感到困惑的数据的操作：

library(caret)
set.seed(3433)
data(AlzheimerDisease)
complete <- data.frame(diagnosis, predictors)
in_train <- createDataPartition(complete$diagnosis, p = 0.75)[[1]]
training <- complete[in_train,]
testing <- complete[-in_train,]
predIL <- grep("^IL", names(training))
smalltrain <- training[, c(1, predIL)]

fit_noPCA <- train(diagnosis ~ ., method = "glm", data = smalltrain)
pre_proc_obj <- preProcess(smalltrain[,-1], method = "pca", thresh = 0.8)
smalltrainsPCs <- predict(pre_proc_obj, smalltrain[,-1])
fit_PCA <- train(x = smalltrainsPCs, y = smalltrain$diagnosis, method = "glm")
fit_noPCA$results$Accuracy
fit_PCA$results$Accuracy

当运行这段代码时，fit_noPCA 的精度为 0.689539，fit_PCA 的精度为 0.682951。但是当我重新运行代码的最后一部分时：

fit_noPCA <- train(diagnosis ~ ., method = "glm", data = smalltrain)
pre_proc_obj <- preProcess(smalltrain[,-1], method = "pca", thresh = 0.8)
smalltrainsPCs <- predict(pre_proc_obj, smalltrain[,-1])
fit_PCA <- train(x = smalltrainsPCs, y = smalltrain$diagnosis, method = "glm")
fit_noPCA$results$Accuracy
fit_PCA$results$Accuracy

然后每次我重新运行这 6 行时，我都会得到不同的准确度值。为什么会这样？是因为我没有重置种子吗？即使，这个过程的固有随机性在哪里？

Answer 1

默认情况下，模型是使用bootstrap训练的，你可以在这里看到：

library(caret)
library(AppliedPredictiveModeling)

> fit_noPCA
Generalized Linear Model 

251 samples
 12 predictor
  2 classes: 'Impaired', 'Control' 

No pre-processing
Resampling: Bootstrapped (25 reps) 
Summary of sample sizes: 251, 251, 251, 251, 251, 251, ... 
Resampling results:

  Accuracy   Kappa     
  0.6870006  0.04107016

所以对于每个 train ，bootstrapped 样本都会不同，为了得到相同的结果，你可以在运行训练之前设置种子：

set.seed(111)
fit_PCA <- train(x = smalltrainsPCs, y = smalltrain$diagnosis, method = "glm",trControl=trainControl(method="boot",number=100))
fit_PCA$results$Accuracy
[1] 0.6983512

set.seed(112)
fit_PCA <- train(x = smalltrainsPCs, y = smalltrain$diagnosis, method = "glm",trControl=trainControl(method="boot",number=100))
fit_PCA$results$Accuracy
[1] 0.6991537

set.seed(111)
fit_PCA <- train(x = smalltrainsPCs, y = smalltrain$diagnosis, method = "glm",trControl=trainControl(method="boot",number=100))
fit_PCA$results$Accuracy
[1] 0.6983512

或者使用例如 cv，您可以在 trainControl

中使用 index= 定义折叠

在 R（插入符号）中重新运行 preProcess()、predict() 和 train() 时模型精度不同

Different model accuracy when rerunning preProcess(), predict() and train() in R (caret)

r

machine-learning

r-caret

glm