Random Forest: OOB performance = test set performance?

Hi ML,

For random forest, I thought that the out-of-bag performance should be
the same (or at least very similar) to the performance calculated on a
separated test set.

But this does not seem to be the case.

In the following code, the accuracy computed on out-of-bag sample is
77.81%, while the one computed on a separated test set is 81%.

Can you please check what I am doing wrong?

Thanks in advance and best regards.

library(randomForest)
library(ISLR)

Carseats$High <- ifelse(Carseats$Sales<=8,"No","Yes")
Carseats$High <- as.factor(Carseats$High)

train = sample(1:nrow(Carseats), 200)

rf = randomForest(High~.-Sales,
                   data=Carseats,
                   subset=train,
                   mtry=6,
                   importance=T)

acc <- (rf$confusion[1,1] + rf$confusion[2,2]) / sum(rf$confusion)
print(paste0("Accuracy OOB: ", round(acc*100,2), "%"))

yhat <- predict(rf, newdata=Carseats[-train,])
y <- Carseats[-train,]$High
conftest <- table(y, yhat)
acctest <- (conftest[1,1] + conftest[2,2]) / sum(conftest)
print(paste0("Accuracy test set: ", round(acctest*100,2), "%"))

I think the only thing you are doing wrong is not setting the random
seed (set.seed()) so your results are not reproducible. Depending on
the random sample used to select the training and test sets, you get
slightly varying accuracy for both, sometimes one is better and
sometimes the other.

HTH,

Peter

On Sat, Apr 10, 2021 at 8:49 PM<thebudget72 at gmail.com>  wrote:

Hi ML,

For random forest, I thought that the out-of-bag performance should be
the same (or at least very similar) to the performance calculated on a
separated test set.

But this does not seem to be the case.

In the following code, the accuracy computed on out-of-bag sample is
77.81%, while the one computed on a separated test set is 81%.

Can you please check what I am doing wrong?

Thanks in advance and best regards.

library(randomForest)
library(ISLR)

Carseats$High <- ifelse(Carseats$Sales<=8,"No","Yes")
Carseats$High <- as.factor(Carseats$High)

train = sample(1:nrow(Carseats), 200)

rf = randomForest(High~.-Sales,
                    data=Carseats,
                    subset=train,
                    mtry=6,
                    importance=T)

acc <- (rf$confusion[1,1] + rf$confusion[2,2]) / sum(rf$confusion)
print(paste0("Accuracy OOB: ", round(acc*100,2), "%"))

yhat <- predict(rf, newdata=Carseats[-train,])
y <- Carseats[-train,]$High
conftest <- table(y, yhat)
acctest <- (conftest[1,1] + conftest[2,2]) / sum(conftest)
print(paste0("Accuracy test set: ", round(acctest*100,2), "%"))