使用分类响应变量的辐射森林预测间隔

Question

我试图从具有分类响应变量的随机森林模型中获取预测间隔。理想情况下，我想看看该模型对将观察结果分类为给定的响应类别有多信心。

在我的代码的最后一行中，您会看到predict（）在不包括interval =参数时可行。当我包含“ Interval =”时，我会出现错误。关于如何获得输出的预测间隔有任何想法吗？

# Load libraries
library(data.table)
library(randomForest)
library(caret)

# Set seed
set.seed(123)

# Load the necessary data
df.0 <- diamonds
setDT(df.0)

# set up the cross-validation parameters
control <- trainControl(method = "repeatedcv",
               number = 10)
metric <- "Accuracy"
mtry <- seq(from = 1,
            to = length(unique(df.0$cut)),
            by = 1)
tunegrid <- expand.grid(mtry = mtry)

# Add rownames so we can use as index
df.0[, indexNum := .I]
trainer <- df.0[ ,.SD[sample(x = .N, size = (.N * 0.9))], by = cut] # Pull 90% of each cut into training
tester <- df.0[!trainer, on = c("indexNum")]

# Remove index number
tester <- tester[, ":=" (indexNum = NULL)] 
trainer <- trainer[, ":=" (indexNum = NULL)] 

# build a model and assess its accuracy via 10-fold cross validation
rf_mod <- 
  train(
    x = trainer[, .(x, y, z, depth, table)],
    y = trainer$cut,
    method = "rf",
    metric = "Accuracy",
    tuneGrid = tunegrid
  )

# check out which mtry value was best
plot(rf_mod)

# test the model against the test data
cut_pred <- predict(rf_mod, newdata = tester[, .(x, y, z, depth, table), interval = "prediction")