test_xgboost_missings.R

library(xgboost)
set.seed(1234)

N = 1000
x1 <- runif(N)
x <- ifelse(x1 <= 0.2, as.numeric(NA), x1)
y <- as.numeric(x1 >= 0.9)

bst <- xgboost(data = matrix(x, ncol=1), label = y,
               objective = "binary:logistic", eval_metric = "logloss",
               nrounds = 1, max_depth = 1, eta = 1., lambda = 0, nthread = 1)

# this shows the two possible leaf values
xgb.dump(model=bst)

pr <- function(xcolumn) {
    md <- matrix(as.numeric(xcolumn), ncol=1)
    ms <- as(md, "dgCMatrix")
    cat('margin from dense : ', predict(bst, xgb.DMatrix(md), outputmargin=T), '\n')
    cat('margin from sparse: ', predict(bst, xgb.DMatrix(ms), outputmargin=T), '\n')
    #print(c('leaf from dense : ', predict(bst, xgb.DMatrix(md), predleaf=T), '\n'))
    #print(c('leaf from sparse: ', predict(bst, xgb.DMatrix(ms), predleaf=T), '\n'))
}

# final row not NA
pr(c(NA,1))
pr(c(1,NA,1)) 

# final row NA
pr(c(NA))
pr(c(NA,NA))
pr(c(NA,NA,NA))