Loop for Repeated K-fold cross validation

Question

I am trying to write a loop for repeated k-fold cross validation. Basically trying to perform a 10-fold cross validation and repeat the process 10-times to get the predictions and the resulting 10 AUC values.

I seem to be missing something in the loop that allows moving the calculated predictions to the respective column of the empty dataframe created for the k-fold results. I only get the last k-fold scores in my output ...instead of all 10. I still have to get the auc values for each k-fold validation.

Is there a way to incorporate the auc calculation in the loop to get the values? Would greatly appreciate if someone can guide me with this.

library(cvTools)
library (glmnet)
#library(pROC) 

k <- 10 #the number of folds
x <- structure(list(PC1 = c(-2.03456672313651, -1.73707505007147, 
-2.03456672313652, -0.255368300655119, -1.73707505007143, -2.03456672313651, 
-0.37500359723752, -2.03456672313651, -2.03456672313651, 3.47288460329945, 
-0.734187869112349, -0.0134149056651377, 0.0942929078885968, 
-2.0345667231365, -2.03456672313651), PC2 = c(0.112471741011579, 
0.133858302549922, 0.1124717410116, 2.61374131070885, 0.133858302549994, 
0.11247174101158, -0.158995891265301, 0.11247174101159, 0.112471741011592, 
-0.260528749768208, -0.503925189558291, 0.194756984230433, 0.318778158034713, 
0.112471741011598, 0.11247174101159), PC3 = c(2.44850389170835, 
2.3403087394181, 2.44850389170835, -2.46949441441314, 2.34030873941815, 
2.44850389170834, 0.123937826076267, 2.44850389170836, 2.44850389170835, 
-0.367483430521022, -0.155846438581532, 0.509441984698824, 0.612816030555617, 
2.44850389170836, 2.44850389170835), PC4 = c(0.112471741011652, 
0.133858302549981, 0.11247174101165, 0.00436673840662417, 0.133858302549995, 
0.112471741011656, -0.158995891265306, 0.112471741011666, 0.112471741011661, 
-0.260528749768211, -0.290253126970872, -2.28110627358792, 0.318778158034689, 
0.11247174101168, 0.11247174101167), PC5 = c(0.112471741011684, 
0.13385830255004, 0.112471741011692, 0.00436673840662224, 0.133858302550053, 
0.112471741011681, -0.158995891265284, 0.112471741011697, 0.112471741011696, 
-0.260528749768212, 1.20999715739728, -1.91404159432553, 0.318778158034758, 
0.112471741011709, 0.112471741011692)), .Names = c("PC1", "PC2", 
"PC3", "PC4", "PC5"), row.names = c("O35245", "O35286", "O54949", 
"O54991", "O88569", "P14733", "P16054", "P21619", "P24369", "P37889", 
"P40201", "P57080", "P60843", "P63085", "P99029"), class = "data.frame")

folds <- cvFolds(NROW(x), K=k)

mypreds <- data.frame(matrix(0, nrow(x),ncol = 10)) # create a dataframe to store results of all 10 k-fold repetititions
row.names(mypreds) <- row.names(x) # row names for the dataframe
names(mypreds) <- paste("K", (1:10), sep = "") # column names

set.seed(123)

j <- 1
nsim = 10 # number of repetitions 

x$kfoldlpred <- rep(0,nrow(x)) # append a column to original dataframe to temporarily store results of each k-fold

# the loop for repeated cross-validation
repeatcv <- function(){
  while (j <= nsim){
    for(i in 1:k){
      train <- x[folds$subsets[folds$which != i], ] #Set the training set 
      train_response <- responseY1[folds$subsets[folds$which != i]] # set the training set response

      validation <- x[folds$subsets[folds$which == i], ] #Set the validation set

      lasso_newglm <- glmnet(as.matrix(train), train_response, alpha = 1,family = "binomial") #Get your new logistic regression model (just fit on the train data)
      lasso_cvglm <- cv.glmnet(as.matrix(train), train_response, alpha = 1, family = "binomial",type.measure = "deviance")
      lasso_newpred <- predict(lasso_newglm,newx = as.matrix(validation), type = "response", s = c(lasso_cvglm$lambda.min)) #Get the predicitons for the validation set (from the model just fit on the train data)

      x[folds$subsets[folds$which == i],]$kfoldlpred <- lasso_newpred
    }
    mypreds[,i] <- x$kfoldlpred
    j <- j+1
  }
  return(mypreds)
}

score 3 · Answer 1 · edited May 23 '17 at 12:18

The caret package provides repeated cross validation out of the box. Here's a minimal working example:

library(caret)
model <- train(x = iris[51:150,1:2], 
               y = factor(iris[51:150,5]), 
               method = 'glmnet', 
               metric='ROC', 
               trControl = trainControl(method = 'repeatedcv', # repeated cross validation
                                        number = 10, # nr of partitions
                                        repeats = 10, # nr of repeats
                                        classProbs = T, 
                                        summaryFunction = twoClassSummary))

model$resample gives you the AUC over all partitions and repeats (with 10 partitions and 10 repeats its 10*10=100 values):

> model$resample
     ROC Sens Spec     Resample
1   0.90  0.8  0.8 Fold05.Rep10
2   0.98  1.0  0.8 Fold04.Rep10
3   0.80  1.0  0.2 Fold01.Rep09
4   0.64  0.4  0.8 Fold08.Rep07
5   0.86  0.8  0.8 Fold05.Rep06
[...]

BTW: if you additionally want to plot the ROC curve over all partitions and repeats, see this question.

Thank you for the reply. I am however trying to self code this as a first step before using the caret package which will be the next step. I will surely use your example as a guide. — DataStudent, Jun 02 '16 at 05:35

score 0 · Answer 2 · answered May 14 '21 at 12:34

0

This is quite late but here is a correction to your code. Replace i with j in the line

mypreds[,i] <- x$kfoldlpred. That's your code should be mypreds[,j] <- x$kfoldlpred

answered May 14 '21 at 12:34

Fonyuy Mercy Nyuybin

1
1

Loop for Repeated K-fold cross validation

2 Answers2