xgboost: Regression

library(mlexperiments)
library(mllrnrs)

See https://github.com/kapsner/mllrnrs/blob/main/R/learner_xgboost.R for implementation details.

Preprocessing

Import and Prepare Data

library(mlbench)
data("BostonHousing")
dataset <- BostonHousing |>
  data.table::as.data.table() |>
  na.omit()

feature_cols <- colnames(dataset)[1:13]
target_col <- "medv"

General Configurations

seed <- 123
if (isTRUE(as.logical(Sys.getenv("_R_CHECK_LIMIT_CORES_")))) {
  # on cran
  ncores <- 2L
} else {
  ncores <- ifelse(
    test = parallel::detectCores() > 4,
    yes = 4L,
    no = ifelse(
      test = parallel::detectCores() < 2L,
      yes = 1L,
      no = parallel::detectCores()
    )
  )
}
options("mlexperiments.bayesian.max_init" = 10L)
options("mlexperiments.optim.xgb.nrounds" = 100L)
options("mlexperiments.optim.xgb.early_stopping_rounds" = 10L)

Generate Training- and Test Data

data_split <- splitTools::partition(
  y = dataset[, get(target_col)],
  p = c(train = 0.7, test = 0.3),
  type = "stratified",
  seed = seed
)

train_x <- model.matrix(
  ~ -1 + .,
  dataset[data_split$train, .SD, .SDcols = feature_cols]
)
train_y <- log(dataset[data_split$train, get(target_col)])


test_x <- model.matrix(
  ~ -1 + .,
  dataset[data_split$test, .SD, .SDcols = feature_cols]
)
test_y <- log(dataset[data_split$test, get(target_col)])

Generate Training Data Folds

fold_list <- splitTools::create_folds(
  y = train_y,
  k = 3,
  type = "stratified",
  seed = seed
)

Experiments

Prepare Experiments

# required learner arguments, not optimized
learner_args <- list(
  objective = "reg:squarederror",
  eval_metric = "rmse"
)

# set arguments for predict function and performance metric,
# required for mlexperiments::MLCrossValidation and
# mlexperiments::MLNestedCV
predict_args <- NULL
performance_metric <- metric("rmsle")
performance_metric_args <- NULL
return_models <- FALSE

# required for grid search and initialization of bayesian optimization
parameter_grid <- expand.grid(
  subsample = seq(0.6, 1, .2),
  colsample_bytree = seq(0.6, 1, .2),
  min_child_weight = seq(1, 5, 4),
  learning_rate = seq(0.1, 0.2, 0.1),
  max_depth = seq(1, 5, 4)
)
# reduce to a maximum of 10 rows
if (nrow(parameter_grid) > 10) {
  set.seed(123)
  sample_rows <- sample(seq_len(nrow(parameter_grid)), 10, FALSE)
  parameter_grid <- kdry::mlh_subset(parameter_grid, sample_rows)
}

# required for bayesian optimization
parameter_bounds <- list(
  subsample = c(0.2, 1),
  colsample_bytree = c(0.2, 1),
  min_child_weight = c(1L, 10L),
  learning_rate = c(0.1, 0.2),
  max_depth =  c(1L, 10L)
)
optim_args <- list(
  iters.n = ncores,
  kappa = 3.5,
  acq = "ucb"
)

Hyperparameter Tuning

Bayesian Optimization

tuner <- mlexperiments::MLTuneParameters$new(
  learner = mllrnrs::LearnerXgboost$new(
    metric_optimization_higher_better = FALSE
  ),
  strategy = "bayesian",
  ncores = ncores,
  seed = seed
)

tuner$parameter_grid <- parameter_grid
tuner$parameter_bounds <- parameter_bounds

tuner$learner_args <- learner_args
tuner$optim_args <- optim_args

tuner$split_type <- "stratified"

tuner$set_data(
  x = train_x,
  y = train_y
)

tuner_results_bayesian <- tuner$execute(k = 3)
#> 
#> Registering parallel backend using 4 cores.

head(tuner_results_bayesian)
#>    Epoch setting_id subsample colsample_bytree min_child_weight learning_rate max_depth gpUtility acqOptimum inBounds Elapsed
#> 1:     0          1       0.6              0.8                5           0.2         1        NA      FALSE     TRUE   1.569
#> 2:     0          2       1.0              0.8                5           0.1         5        NA      FALSE     TRUE   1.663
#> 3:     0          3       0.8              0.8                5           0.1         1        NA      FALSE     TRUE   1.611
#> 4:     0          4       0.6              0.8                5           0.2         5        NA      FALSE     TRUE   1.611
#> 5:     0          5       1.0              0.8                1           0.1         5        NA      FALSE     TRUE   0.941
#> 6:     0          6       0.8              0.8                5           0.1         5        NA      FALSE     TRUE   0.906
#>         Score metric_optim_mean nrounds errorMessage        objective eval_metric
#> 1: -0.1865024         0.1865024      56           NA reg:squarederror        rmse
#> 2: -0.1607242         0.1607242      89           NA reg:squarederror        rmse
#> 3: -0.1913163         0.1913163     100           NA reg:squarederror        rmse
#> 4: -0.1609879         0.1609879      66           NA reg:squarederror        rmse
#> 5: -0.1573682         0.1573682     100           NA reg:squarederror        rmse
#> 6: -0.1635603         0.1635603      92           NA reg:squarederror        rmse

k-Fold Cross Validation

validator <- mlexperiments::MLCrossValidation$new(
  learner = mllrnrs::LearnerXgboost$new(
    metric_optimization_higher_better = FALSE
  ),
  fold_list = fold_list,
  ncores = ncores,
  seed = seed
)

validator$learner_args <- tuner$results$best.setting[-1]

validator$predict_args <- predict_args
validator$performance_metric <- performance_metric
validator$performance_metric_args <- performance_metric_args
validator$return_models <- return_models

validator$set_data(
  x = train_x,
  y = train_y
)

validator_results <- validator$execute()
#> 
#> CV fold: Fold1
#> 
#> CV fold: Fold2
#> 
#> CV fold: Fold3

head(validator_results)
#>     fold performance subsample colsample_bytree min_child_weight learning_rate max_depth nrounds        objective eval_metric
#> 1: Fold1  0.04193925       0.6                1                1           0.1         5      92 reg:squarederror        rmse
#> 2: Fold2  0.05079392       0.6                1                1           0.1         5      92 reg:squarederror        rmse
#> 3: Fold3  0.03915493       0.6                1                1           0.1         5      92 reg:squarederror        rmse

Nested Cross Validation

Inner Bayesian Optimization

validator <- mlexperiments::MLNestedCV$new(
  learner = mllrnrs::LearnerXgboost$new(
    metric_optimization_higher_better = FALSE
  ),
  strategy = "bayesian",
  fold_list = fold_list,
  k_tuning = 3L,
  ncores = ncores,
  seed = seed
)

validator$parameter_grid <- parameter_grid
validator$learner_args <- learner_args
validator$split_type <- "stratified"


validator$parameter_bounds <- parameter_bounds
validator$optim_args <- optim_args

validator$predict_args <- predict_args
validator$performance_metric <- performance_metric
validator$performance_metric_args <- performance_metric_args
validator$return_models <- TRUE

validator$set_data(
  x = train_x,
  y = train_y
)

validator_results <- validator$execute()
#> 
#> CV fold: Fold1
#> 
#> Registering parallel backend using 4 cores.
#> 
#> CV fold: Fold2
#> CV progress [====================================================================>-----------------------------------] 2/3 ( 67%)
#> 
#> Registering parallel backend using 4 cores.
#> 
#> CV fold: Fold3
#> CV progress [========================================================================================================] 3/3 (100%)
#>                                                                                                                                   
#> Registering parallel backend using 4 cores.

head(validator_results)
#>     fold performance subsample colsample_bytree min_child_weight learning_rate max_depth nrounds        objective eval_metric
#> 1: Fold1  0.04147964 0.6225939        0.9208933                5     0.1326066         5      59 reg:squarederror        rmse
#> 2: Fold2  0.05881907 1.0000000        0.8000000                1     0.1000000         5      94 reg:squarederror        rmse
#> 3: Fold3  0.03890190 0.6000000        1.0000000                5     0.2000000         5      37 reg:squarederror        rmse

Holdout Test Dataset Performance

Predict Outcome in Holdout Test Dataset

preds_xgboost <- mlexperiments::predictions(
  object = validator,
  newdata = test_x
)

Evaluate Performance on Holdout Test Dataset

perf_xgboost <- mlexperiments::performance(
  object = validator,
  prediction_results = preds_xgboost,
  y_ground_truth = test_y,
  type = "regression"
)
perf_xgboost
#>    model performance        mse        msle       mae       mape      rmse      rmsle       rsq      sse
#> 1: Fold1  0.04322328 0.02725729 0.001868252 0.1188479 0.04074989 0.1650978 0.04322328 0.8227146 4.224880
#> 2: Fold2  0.04730978 0.03081692 0.002238216 0.1235033 0.04247960 0.1755475 0.04730978 0.7995622 4.776623
#> 3: Fold3  0.03977942 0.02204549 0.001582402 0.1090010 0.03781531 0.1484773 0.03977942 0.8566129 3.417052