The extracted gbm final model does not return the same result as the trained gbm model

Question

I'm trying to use the final model extracted from the trained gbm model, but the extracted model does not return the factorized result as the trained model does. It seems that the extracted final model works according to the returned value, however, it just returns the values of calculation. How can I get the factorized result as the trained model.

library(caret)
library(mlbench)

data(Sonar)
set.seed(7)

Sonar$Class <- ifelse(Sonar$Class == 'R', 0, 1)
Sonar$Class <- as.factor(Sonar$Class)
validation_index <- createDataPartition(Sonar$Class, p=0.80, list=FALSE)
validation <- Sonar[-validation_index,]
training <- Sonar[validation_index,]
outcomename <- 'Class'
predictors <- names(training)[!names(training) %in% outcomename]

set.seed(7)
control <- trainControl(method = "repeatedcv",  number = 5,  repeats = 5)
model_gbm <- train(training[, predictors], training[, outcomename], method = 'gbm', trControl = control, tuneLength = 10)

predict(model_gbm, validation[,1:60])
[1] 1 1 0 1 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
Levels: 0 1

predict(model_gbm$finalModel, validation[,1:60], n.trees = 300)
[1]  -3.1174531  -1.8335718   5.0780422  -8.6681791   8.9634393  -1.4079936  11.7232458
[8]  18.4189859  14.3978772  11.3605253  13.4694812  10.2752696  11.4957672  10.0370462
[15]   8.6009983   0.3718381   0.1297673   2.4099186   6.7774090 -10.8356795 -10.1842065
[22]  -2.3222431  -8.1525336  -3.3665867 -10.7953353  -2.4607156 -11.4277641  -4.7164270
[29]  -6.3882544  -3.7306579  -6.9323133  -4.2643347  -0.2128462  -9.3395850 -13.0759289
[36] -12.8259643  -6.5314340 -12.7968160 -16.6217507 -12.0370978  -3.1100361

Answer 1

The predict.gbm function has a type argument, which can be "response" or "link". To get the predicted probabilities one should set it to "response". Then to convert these predictions to a class one can use a threshold value (0.5 is used by caret train). To get and idea here is an example:

library(caret)
library(mlbench)

data(Sonar)
set.seed(7)

validation_index <- createDataPartition(Sonar$Class, p=0.80, list=FALSE)
validation <- Sonar[-validation_index,]
training <- Sonar[validation_index,]

set.seed(7)
control <- trainControl(method = "repeatedcv",
                        number = 2,
                        repeats = 2)
model_gbm <- train(Class~.,
                   data = training,
                   method = 'gbm',
                   trControl = control,
                   tuneLength = 3)

predict using caret:

preds1 <- predict(model_gbm, validation[,1:60], type = "prob")

predict using gbm:

library(gbm)
preds2 <- predict(model_gbm$finalModel, validation[,1:60], n.trees = 100, type = "response")

all.equal(preds1[,1], preds2)
#output
TRUE

or in case of classes:

preds1_class <- predict(model_gbm, validation[,1:60])

to check if they are equal to gbm predictions threshold the predictions:

all.equal(
  as.factor(ifelse(preds2 > 0.5, "M", "R")),
  preds1_class)
#output
TRUE