xgboost sending sparse zeroes to missing node

xgboost_sparse_zeroes.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))
  )
}

N <- 1000

# Need a seed that will cause zero to get sent to the non-contiguous node
set.seed(1236)

train <- generate_data(N)

smm_train <- sparse.model.matrix(response ~ int1, train)
dtrain <- xgb.DMatrix(data = smm_train, label = train[, response])

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)

xgb.dump(model = model)
#   booster[0]
#   0:[f1<6.5] yes=1,no=2,missing=2" "1:leaf=-1.11714
#   2:leaf=-1.37056
# -> Split int1 at 6.5

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

train[, mean(pred), by = int1][order(int1)]
#     int1        V1
#  1:   -6 0.2465423
#  2:   -5 0.2465423
#  3:   -4 0.2465423
#  4:   -3 0.2465423
#  5:   -2 0.2465423
#  6:   -1 0.2465423
#  7:    0 0.2025297    <- !
#  8:    1 0.2465423
#  9:    2 0.2465423
# 10:    3 0.2465423
# 11:    4 0.2465423
# 12:    5 0.2465423
# 13:    6 0.2465423
# 14:    7 0.2025297
# 15:    8 0.2025297
# 16:    9 0.2025297
# 17:   10 0.2025297
# 18:   11 0.2025297
# 19:   12 0.2025297
# 20:   13 0.2025297