lightgbm: Regression

library(mlexperiments)
library(mllrnrs)

See https://github.com/kapsner/mllrnrs/blob/main/R/learner_lightgbm.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.lgb.nrounds" = 100L)
options("mlexperiments.optim.lgb.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$train, .SD, .SDcols = feature_cols]
)
train_y <- dataset[data_split$train, get(target_col)]


test_x <- model.matrix(
  ~ -1 + .,
  dataset[data_split$test, .SD, .SDcols = feature_cols]
)
test_y <- dataset[data_split$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(
  max_depth = -1L,
  verbose = -1L,
  objective = "regression",
  metric = "l2"
)

# 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(
  bagging_fraction = seq(0.6, 1, .2),
  feature_fraction = seq(0.6, 1, .2),
  min_data_in_leaf = seq(2, 10, 2),
  learning_rate = seq(0.1, 0.2, 0.1),
  num_leaves = seq(2, 20, 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(
  bagging_fraction = c(0.2, 1),
  feature_fraction = c(0.2, 1),
  min_data_in_leaf = c(2L, 12L),
  learning_rate = c(0.1, 0.2),
  num_leaves =  c(2L, 20L)
)
optim_args <- list(
  iters.n = ncores,
  kappa = 3.5,
  acq = "ucb"
)

Hyperparameter Tuning

Bayesian Optimization

tuner <- mlexperiments::MLTuneParameters$new(
  learner = mllrnrs::LearnerLightgbm$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 bagging_fraction feature_fraction min_data_in_leaf learning_rate num_leaves gpUtility acqOptimum inBounds
#> 1:     0          1              0.6              0.6                4           0.2         18        NA      FALSE     TRUE
#> 2:     0          2              0.8              1.0               10           0.2          6        NA      FALSE     TRUE
#> 3:     0          3              0.8              0.8                4           0.1          2        NA      FALSE     TRUE
#> 4:     0          4              1.0              0.8                4           0.1         10        NA      FALSE     TRUE
#> 5:     0          5              1.0              0.6                6           0.2         18        NA      FALSE     TRUE
#> 6:     0          6              1.0              1.0                8           0.1         14        NA      FALSE     TRUE
#>    Elapsed     Score metric_optim_mean nrounds errorMessage max_depth verbose  objective metric
#> 1:   1.081 -13.59085          13.59085      85           NA        -1      -1 regression     l2
#> 2:   1.072 -13.75483          13.75483      55           NA        -1      -1 regression     l2
#> 3:   1.044 -21.08526          21.08526      58           NA        -1      -1 regression     l2
#> 4:   1.126 -13.31343          13.31343      92           NA        -1      -1 regression     l2
#> 5:   0.104 -13.86649          13.86649      80           NA        -1      -1 regression     l2
#> 6:   0.106 -14.58646          14.58646     100           NA        -1      -1 regression     l2

k-Fold Cross Validation

validator <- mlexperiments::MLCrossValidation$new(
  learner = mllrnrs::LearnerLightgbm$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 bagging_fraction feature_fraction min_data_in_leaf learning_rate num_leaves nrounds max_depth verbose
#> 1: Fold1   0.1572748              0.6              0.8                2           0.2         10      34        -1      -1
#> 2: Fold2   0.1770563              0.6              0.8                2           0.2         10      34        -1      -1
#> 3: Fold3   0.1439331              0.6              0.8                2           0.2         10      34        -1      -1
#>     objective metric
#> 1: regression     l2
#> 2: regression     l2
#> 3: regression     l2

Nested Cross Validation

Inner Bayesian Optimization

validator <- mlexperiments::MLNestedCV$new(
  learner = mllrnrs::LearnerLightgbm$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 bagging_fraction feature_fraction min_data_in_leaf learning_rate num_leaves nrounds max_depth verbose
#> 1: Fold1   0.1972673        1.0000000        0.2000000                7     0.1000000          2     100        -1      -1
#> 2: Fold2   0.1934754        0.5029800        0.4977050                7     0.1195995          4      52        -1      -1
#> 3: Fold3   0.1391255        0.8050493        0.5902201                2     0.1458152         20      44        -1      -1
#>     objective metric
#> 1: regression     l2
#> 2: regression     l2
#> 3: regression     l2

Holdout Test Dataset Performance

Predict Outcome in Holdout Test Dataset

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

Evaluate Performance on Holdout Test Dataset

perf_lightgbm <- mlexperiments::performance(
  object = validator,
  prediction_results = preds_lightgbm,
  y_ground_truth = test_y,
  type = "regression"
)
perf_lightgbm
#>    model performance       mse       msle      mae      mape     rmse     rmsle       rsq      sse
#> 1: Fold1   0.1593611 16.258140 0.02539596 2.700096 0.1230206 4.032138 0.1593611 0.7945966 2520.012
#> 2: Fold2   0.1720557 12.614833 0.02960318 2.629003 0.1329465 3.551737 0.1720557 0.8406257 1955.299
#> 3: Fold3   0.1424064  9.666178 0.02027958 2.226562 0.1101779 3.109048 0.1424064 0.8778787 1498.258