LightGBM for general utility

lightgbm_pmse.R

library(synthpop)
library(lightgbm)
library(tidyverse)

df <- SD2011

syn_res <- syn(df[,1:7])
sdf <- syn_res$syn


train_params <- list(
  learning_rate = 1.0, 
  objective = "binary",
  nthread = 6L,
  force_row_wise = TRUE
)

pmse_dat <- bind_rows(df[,1:7], sdf) |> mutate(label = rep(c(0L, 1L), each = nrow(df)))
y <- model.response(model.frame(label ~ ., data = pmse_dat))
X <- model.matrix(label ~ ., data = pmse_dat)
Xs <- model.matrix(label ~ ., data = pmse_dat |> filter(label == 1))

lgbdata <- lgb.Dataset(data = unname(X), label = y)

n_sim <- 20L
max_nrounds <- 10000L
Pmat <- matrix(0.0, nrow(Xs), n_sim)
round_seq <- round(seq(1, max_nrounds, length.out = n_sim))
for (i in seq_along(round_seq)) {
  cat("iteration", i, "\n")
  fit <- lgb.train(params = train_params, data = lgbdata, nrounds = round_seq[i], 
                   verbose = 0L)
  Pmat[,i] <- predict(fit, Xs)
}

pMSEs <- colMeans((Pmat - nrow(Xs) / nrow(X))^2)
plot(round_seq, pMSEs, type = "l")