Cross-validation error

[bgreenwell]

# Load required packages
library(gbm)

# Load the data
data(spam, package = "kernlab") 

# Fit a GBM
set.seed(1913)  # for reproducibility
spam_gbm <- gbm(
  ifelse(type == "spam", 1, 0) ~ ., 
  data = spam, 
  distribution = "bernoulli",
  n.trees = 10,
  interaction.depth = 5,
  shrinkage = 0.001,
  bag.fraction = 1,
  train.fraction = 0.7,
  cv.folds = 5,  #<< causes error
  verbose = TRUE
)

# Error in object$var.levels[[i]] : subscript out of bounds

[bgreenwell]

Works fine when the response is coded outside of gbm():

spam$type <- ifelse(spam$type == "spam", 1, 0)
spam_gbm <- gbm(
  type ~ ., 
  data = spam, 
  distribution = "bernoulli",
  n.trees = 10,
  interaction.depth = 5,
  shrinkage = 0.001,
  bag.fraction = 1,
  train.fraction = 0.7,
  cv.folds = 5,  
  verbose = TRUE
)

[bgreenwell]

Looks like the error occurs in gbmCrossValPredictions(); see below. The commented out line would cause the actual response column to be returned in the prediction data, which predict.gbm() filters):

 # my.data  <- data[flag, !(data.names %in% model$response.name)]
 my.data  <- data[flag, model$var.names]  # patch