如何使用插入符号包模型结果对概率运行 gbm 模拟

Question

I am not sure if this is the right approach, and I'd be happy to be corrected.

Fitting the model

I fitted a gbm model using the caret package in R environment, for sake of an example I will do it without any parameter adjustment and using the iris dataset:

library(caret)
data(iris)
gbmFit <- train(Species ~ ., data = iris, method = "gbm")

This allows me to classify to what species a flower would belong if I had the 4 measurements in the iris dataset.

I am interested in using the results of the predict function in the type = "prob" to run simulations.

Using predict to get the probabilities of each species

Since this is just an example, I don't have new data, so I will use the same data as if it was new, I used the predict function to get the probabilities that each flower is of one species:

PROBS <- predict(gbmFit4, iris[,1:4], type="prob")

this are examples of the result

head(PROBS)
     setosa   versicolor    virginica
1 0.9999989 1.087268e-06 1.679813e-10
2 0.9999998 1.689137e-07 1.404242e-09
3 0.9999995 5.381312e-07 3.131823e-10
4 0.9999996 4.335414e-07 3.912857e-10
5 0.9999989 1.087268e-06 1.679813e-10
6 0.9999987 1.278968e-06 1.679813e-10

Running a simulation for one flower

I know how to do a simulation for one flower, where I use the PROB dataframe to give me the probabilities of a flower to be of each species, and then use sample to simulate the classification given that probability, to make, lets say 1000000 classifications, I use row 107 because it is a less certain case:

set.seed(123)
summary(as.factor(sample(c("setosa", "versicolor", "virginica"), size = 100000, replace = TRUE, prob = PROBS[107,])))

which results in

versicolor  virginica 
     14731      85269 

How to do it for every observation

My goal is to run a simulation like this with new data and get the following result: in average how many flowers were classified in each species for each simulation (av_class_species); and what was the minimum and max for each species (min_class_species, max_class_species), as an example I made this fake data frame (there are 150 flowers in the dataset):

av_class_setosa  max_class_setosa min_class_setosa av_class_versicolor...
       24.4         35                 12                30.2

any help would be greatly appreciated

Answer 1

Found my own answer, although I would be happy if someone found a more efficient way, this is for running 100 simulations:

SIMUL <- list()

for(i in 1:100){
  species <- list()
  for(j in 1:nrow(PROBS)){
    species[[j]] <- sample(c("setosa", "versicolor", "virginica"), size = 1, replace = TRUE, prob = PROBS[j,])
  }
  SIMUL[[i]] <- as.data.frame(table(unlist(species)))
}

SIMUL <- do.call("rbind", SIMUL)

SIMUL <- dplyr::group_by(SIMUL, Var1)
SIMUL <- dplyr::summarise(SIMUL, MEAN_class = mean(Freq), MIN_Class = min(Freq), MAX_Class = max(Freq))

This will result in:

SIMUL

Source: local data frame [3 x 4]

        Var1 MEAN_class MIN_Class MAX_Class
      (fctr)      (dbl)     (int)     (int)
1     setosa       50.0        50        50
2 versicolor       49.7        47        53
3  virginica       50.3        47        53