xgboost_what_are_you_doing.R

library("data.table")
library("xgboost")
library("Matrix")


generate_data <- function(N) {
  data.table(
    response = as.numeric(runif(N) > 0.8),
    int1 = round(rnorm(N, 3, 3)),
    int2 = round(rnorm(N, 3, 3)),
    int3 = round(rnorm(N, 3, 3))
  )
}


N <- 1000
set.seed(1235)  # This seed splits at a point that reproduces bug

mform <- as.formula(response ~ int1 + int2 + int3)

train <- generate_data(N)

lbl_train <- train[, response]
smm_train <- sparse.model.matrix(mform, train)
dtrain <- xgb.DMatrix(data = smm_train, label = lbl_train)

model <- xgb.train(params = list(eta = 1,                
                                 max_depth = 1,            
                                 min_child_weight = 10,    
                                 subsample = 1.0,          
                                 objective = "binary:logistic",
                                 eval_metric = "logloss"),
                    data = dtrain,
                    nrounds = 1,                             
                    nthread = 1,
                    verbose = 1,
                    print.every.n = 1,
                    save_period = 0,
                    save_name = "xgboost.model")

xgb.dump(model = model)
# -> Split f3 (int3) at 3.5

train[, pred := predict(model, dtrain)]

train[, mean(pred), by = int3][order(int3)]
#     int3        V1
#  1:   -6 0.2593045
#  2:   -5 0.2593045
#  3:   -4 0.2593045
#  4:   -3 0.2593045
#  5:   -2 0.2593045
#  6:   -1 0.2593045
#  7:    0 0.2197411    <- whoa dude!
#  8:    1 0.2593045
#  9:    2 0.2593045
# 10:    3 0.2593045
# 11:    4 0.2197411
# 12:    5 0.2197411
# 13:    6 0.2197411
# 14:    7 0.2197411
# 15:    8 0.2197411
# 16:    9 0.2197411
# 17:   10 0.2197411
# 18:   11 0.2197411
# 19:   12 0.2197411
# 20:   15 0.2197411