R caret held-out sample and testing set ROC - r

Question

I am building two different classifiers to predict a binary out come. Then I want to compare the results of the two models by using a ROC curve and the area under it (AUC).
I split the data set into a training and testing set. On the training set I perform a form of cross-validation. From the held-out samples of the cross validation I am able to build a ROC curve per model. Then I use the models on the testing set and build another set of ROC curves.
The results are contradictory which is confusing me. I am not sure which result is the correct one or if I am doing something completely wrong. The held-out sample ROC curve shows that RF is the better model and the training set ROC curve shows that SVM is the better model.
Analysis
library(ggplot2)
library(caret)
library(pROC)
library(ggthemes)
library(plyr)
library(ROCR)
library(reshape2)
library(gridExtra)
my_data <- read.csv("http://www.ats.ucla.edu/stat/data/binary.csv")
str(my_data)
names(my_data)[1] <- "Class"
my_data$Class <- ifelse(my_data$Class == 1, "event", "noevent")
my_data$Class <- factor(emr$Class, levels = c("noevent", "event"), ordered = TRUE)
set.seed(1732)
ind <- createDataPartition(my_data$Class, p = 2/3, list = FALSE)
train <- my_data[ ind,]
test <- my_data[-ind,]
Next I train two models: Random Forest and SVM. Here I also use Max Kuhns function to get the averaged ROC curves from held-out samples for both models and save those results into a another data.frame along with the AUC from the curves.
#Train RF
ctrl <- trainControl(method = "repeatedcv",
number = 5,
repeats = 3,
classProbs = TRUE,
savePredictions = TRUE,
summaryFunction = twoClassSummary)
grid <- data.frame(mtry = seq(1,3,1))
set.seed(1537)
rf_mod <- train(Class ~ .,
data = train,
method = "rf",
metric = "ROC",
tuneGrid = grid,
ntree = 1000,
trControl = ctrl)
rfClasses <- predict(rf_mod, test)
#This is the ROC curve from held out samples. Source is from Max Kuhns 2016 UseR! code here: https://github.com/topepo/useR2016
roc_train <- function(object, best_only = TRUE, ...) {
lvs <- object$modelInfo$levels(object$finalModel)
if(best_only) {
object$pred <- merge(object$pred, object$bestTune)
}
## find tuning parameter names
p_names <- as.character(object$modelInfo$parameters$parameter)
p_combos <- object$pred[, p_names, drop = FALSE]
## average probabilities across resamples
object$pred <- plyr::ddply(.data = object$pred,
.variables = c("obs", "rowIndex", p_names),
.fun = function(dat, lvls = lvs) {
out <- mean(dat[, lvls[1]])
names(out) <- lvls[1]
out
})
make_roc <- function(x, lvls = lvs, nms = NULL, ...) {
out <- pROC::roc(response = x$obs,
predictor = x[, lvls[1]],
levels = rev(lvls))
out$model_param <- x[1,nms,drop = FALSE]
out
}
out <- plyr::dlply(.data = object$pred,
.variables = p_names,
.fun = make_roc,
lvls = lvs,
nms = p_names)
if(length(out) == 1) out <- out[[1]]
out
}
temp <- roc_train(rf_mod)
plot_data_ROC <- data.frame(Model='Random Forest', sens = temp$sensitivities, spec=1-temp$specificities)
#This is the AUC of the held-out samples roc curve for RF
auc.1 <- abs(sum(diff(1-temp$specificities) * (head(temp$sensitivities,-1)+tail(temp$sensitivities,-1)))/2)
#Build SVM
set.seed(1537)
svm_mod <- train(Class ~ .,
data = train,
method = "svmRadial",
metric = "ROC",
trControl = ctrl)
svmClasses <- predict(svm_mod, test)
#ROC curve into df
temp <- roc_train(svm_mod)
plot_data_ROC <- rbind(plot_data_ROC, data.frame(Model='Support Vector Machine', sens = temp$sensitivities, spec=1-temp$specificities))
#This is the AUC of the held-out samples roc curve for SVM
auc.2 <- abs(sum(diff(1-temp$specificities) * (head(temp$sensitivities,-1)+tail(temp$sensitivities,-1)))/2)
Next I will plot the results
#Plotting Final
#ROC of held-out samples
q <- ggplot(data=plot_data_ROC, aes(x=spec, y=sens, group = Model, colour = Model))
q <- q + geom_path() + geom_abline(intercept = 0, slope = 1) + xlab("False Positive Rate (1-Specificity)") + ylab("True Positive Rate (Sensitivity)")
q + theme(axis.line = element_line(), axis.text=element_text(color='black'),
axis.title = element_text(colour = 'black'), legend.text=element_text(), legend.title=element_text())
#ROC of testing set
rf.probs <- predict(rf_mod, test,type="prob")
pr <- prediction(rf.probs$event, factor(test$Class, levels = c("noevent", "event"), ordered = TRUE))
pe <- performance(pr, "tpr", "fpr")
roc.data <- data.frame(Model='Random Forest',fpr=unlist(pe#x.values), tpr=unlist(pe#y.values))
svm.probs <- predict(svm_mod, test,type="prob")
pr <- prediction(svm.probs$event, factor(test$Class, levels = c("noevent", "event"), ordered = TRUE))
pe <- performance(pr, "tpr", "fpr")
roc.data <- rbind(roc.data, data.frame(Model='Support Vector Machine',fpr=unlist(pe#x.values), tpr=unlist(pe#y.values)))
q <- ggplot(data=roc.data, aes(x=fpr, y=tpr, group = Model, colour = Model))
q <- q + geom_line() + geom_abline(intercept = 0, slope = 1) + xlab("False Positive Rate (1-Specificity)") + ylab("True Positive Rate (Sensitivity)")
q + theme(axis.line = element_line(), axis.text=element_text(color='black'),
axis.title = element_text(colour = 'black'), legend.text=element_text(), legend.title=element_text())
#AUC of hold out samples
data.frame(Rf = auc.1, Svm = auc.2)
#AUC of testing set. Source is from Max Kuhns 2016 UseR! code here: https://github.com/topepo/useR2016
test_pred <- data.frame(Class = factor(test$Class, levels = c("noevent", "event"), ordered = TRUE))
test_pred$Rf <- predict(rf_mod, test, type = "prob")[, "event"]
test_pred$Svm <- predict(svm_mod, test, type = "prob")[, "event"]
get_auc <- function(pred, ref){
auc(roc(ref, pred, levels = rev(levels(ref))))
}
apply(test_pred[, -1], 2, get_auc, ref = test_pred$Class)
The results from the held-out samples and from the testing set are totally different (I know they will be different but by this much?).
Rf Svm
0.656044 0.5983193
Rf Svm
0.6326531 0.6453428
From the held-out samples one would choose the RF model but from the testing set one would pick the SVM model.
Which is the "correct" or "better" way to chose the model?
Am I making a big mistake somewhere or not understanding something correctly?

Answer 1

If I understand correctly then you have 3 labeled data sets:
Training
Hold-out CV sample from training
"Testing" CV sample
While, yes, under a hold-out sample CV strategy you normally choose your model based on the hold-out sample, you also don't normally also have a larger validation data sample.
Clearly, if both the hold-out and the Testing data sets are (a) labeled and (b) as close to the level of orthogonality as possible from from the training data, then you'd choose your model based on whichever has the larger sample size.
In your case it looks like what you're calling the hold-out sample is just the repeated CV resampling from training. That being the case you have even more reason to prefer the results from the Testing data set validation. See Steffen's related note on repeated CV.
In theory Random Forest's bagging has a inherit form of cross-validation through the OOB stats and the CV conducted within the training phase should give you some measure of validation. However, in practice it's common to observe a lack of orthogonality and an increased likelihood of overfitting since the samples are coming from the training data itself and may be reinforcing the mistake of overfitting for accuracy.
I can explain that theoretically as above to some extent, then beyond that I just have to tell you that empirically I've found that the performance results from the so-called CV and OOB error calculated from the training data can be highly misleading and the true hold-out (Testing) data that was never touched during training is the far better validation.
Your true hold-out sample is the Testing data set, since none of its data is using during training. Use those results.