xgboost creating different model structures based on choice of sparsity

xgboost_sparse_dense_different_model.R

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


generate_data <- function(N) {
  data.table(
    response = as.numeric(runif(N) > 0.8),
    float1   = rnorm(N, 3, 3))
}

N <- 1000

set.seed(1236)

train <- generate_data(N)

# Set one row to exactly zero or else nothing weird happens
# Comment out this line and model structure is the same for sparse/dense
train[1, float1 := 0]

mm_train <- model.matrix(response ~ float1, train)
smm_train <- sparse.model.matrix(response ~ float1, train)

dtrain <- xgb.DMatrix(data = mm_train, label = train[, response])
dtrain_sparse <- 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)

model_sparse <- xgb.train(params = list(eta = 1,
                                        max_depth = 1,
                                        min_child_weight = 10,
                                        subsample = 1.0,
                                        objective = "binary:logistic",
                                        eval_metric = "logloss"),
                           data = dtrain_sparse,
                           nrounds = 1)

# Different leaf predictions for sparse and dense encodings
xgb.dump(model = model)
# [1] "booster[0]"                         "0:[f1<6.8692] yes=1,no=2,missing=1"
# [3] "1:leaf=-1.12623"                    "2:leaf=-1.54639"                   

xgb.dump(model = model_sparse)
# [1] "booster[0]"                         "0:[f1<6.8692] yes=1,no=2,missing=2"
# [3] "1:leaf=-1.12527"                    "2:leaf=-1.55102"        

sessionInfo()
# R version 3.2.2 (2015-08-14)
# Platform: x86_64-pc-linux-gnu (64-bit)
# Running under: Ubuntu 15.10
# 
# locale:
#  [1] LC_CTYPE=en_IE.UTF-8       LC_NUMERIC=C              
#  [3] LC_TIME=en_IE.UTF-8        LC_COLLATE=en_IE.UTF-8    
#  [5] LC_MONETARY=en_IE.UTF-8    LC_MESSAGES=en_IE.UTF-8   
#  [7] LC_PAPER=en_IE.UTF-8       LC_NAME=C                 
#  [9] LC_ADDRESS=C               LC_TELEPHONE=C            
# [11] LC_MEASUREMENT=en_IE.UTF-8 LC_IDENTIFICATION=C       
# 
# attached base packages:
# [1] stats     graphics  grDevices utils     datasets  methods   base     
# 
# other attached packages:
# [1] Matrix_1.2-2     xgboost_0.4-3    setwidth_1.0-4   colorout_1.1-1  
# [5] magrittr_1.5     data.table_1.9.6 devtools_1.9.1  
# 
# loaded via a namespace (and not attached):
# [1] tools_3.2.2     memoise_0.2.1   stringi_1.0-1   grid_3.2.2     
# [5] stringr_1.0.0   digest_0.6.8    chron_2.3-47    lattice_0.20-33