mlexperiments: Getting Started

The goal of the package mlexperiments is to provide an extensible framework for reproducible machine learning experiments, namely:

  • Hyperparameter tuning: with the R6 class mlexperiments::MLTuneParameters, to optimize the hyperparameters in a k-fold cross-validation with one of the two strategies
  • K-fold Cross-validation (CV): with the R6 class mlexperiments::MLCrossValidation, to validate one hyperparameter setting
  • Nested k-fold cross validation: with the R6 class mlexperiments::MLNestedCV, which basically combines the two experiments above to perform a hyperparameter optimization on an inner CV loop, and to validate the best hyperparameter setting on an outer CV loop

The package provides a minimal shell for these experiments, and - with few adjustments - users can prepare different learner algorithms so that they can be used with mlexperiments.

This vignette will go through the steps that are necessary to prepare a new learner.

General Overview

In general, the learner class exposes 4 methods that can be defined:

  • $fit A wrapper around the private function fun_fit, which needs to be defined for every learner. The return value of this function is the fitted model.
  • $predict A wrapper around the private function fun_predict, which needs to be defined for every learner. The function must accept the three arguments model, newdata, and ncores and is a wrapper around the respective learner’s predict-function. In order to allow the passing of further arguments, the ellipsis (...) can be used. The function should return the prediction results.
  • $cross_validation A wrapper around the private function fun_optim_cv, which needs to be defined when hyperparameters should be optimized with a grid search (required for use with mlexperiments::MLTuneParameters, and mlexperiments::MLNestedCV).
  • $bayesian_scoring_function A wrapper around the private function fun_bayesian_scoring_function, which needs to be defined when hyperparameters should be optimized with a Bayesian process (required for use with mlexperiments::MLTuneParameters, and mlexperiments::MLNestedCV).

In the following, we will go through the steps to prepare the algorithm ‘class::knn()’ to be used with mlexperiments (the same code is also implemented in the package and ready to use as mlexperiments::LearnerKnn).

Steps to Prepare an Algorithm for Use with mlexperiments

The fit Method

This method must take the arguments x, y, ncores, seed, as well as the ellipsis (...), while arguments to parameterize the learner are to be passed to the function with the latter. The fit method should include one call to fit a model of the algorithm and it should finally return the fitted model.

knn_fit <- function(x, y, ncores, seed, ...) {
  kwargs <- list(...)
  stopifnot("k" %in% names(kwargs))
  args <- kdry::list.append(list(train = x, cl = y), kwargs)
  args$prob <- TRUE
  set.seed(seed)
  fit <- do.call(class::knn, args)
  return(fit)
}

The predict Method

This method must take the arguments model, newdata, ncores, and the ellipsis (...). It is a wrapper around the respective algorithm’s predict() function, while specific arguments required to parameterize it can be passed with the ellipsis. The experiments mlexperiments::MLCrossValidation and mlexperiments::MLNestedCV do both have the field $predict_args to define a list that is further passed on to the predcit method’s ellipsis. In contrast, when it is required to further parameterize this method during the hyperparameter tuning (mlexperiments::MLTuneParameters), it is required to define those parameters within the cross_validation method (see below). The returned value of the predict method should be a vector with the predictions.

knn_predict <- function(model, newdata, ncores, ...) {
  kwargs <- list(...)
  stopifnot("type" %in% names(kwargs))
  if (kwargs$type == "response") {
    return(model)
  } else if (kwargs$type == "prob") {
    # there is no knn-model but the probabilities predicted for the test data
    return(attributes(model)$prob)
  }
}

The implementation of class::knn() is in some ways special and different from the implementation of other algorithms. One of these peculiarities is that class::knn() does not return a fitted model but instead returns the predicted values directly. Depending on the value of the argument prob, these results also include the probability values of the predicted classes.

The cross_validation Method

The purpose of this function is to perform a k-fold cross validation for one specific hyperparameter setting. The function must take the arguments x, y, params (a list of hyperparameters), fold_list (to define the cross-validation folds), ncores, and seed. Finally, the function must return a named list with at least one item called metric_optim_mean, which contains the cross validated error metric.

knn_optimization <- function(x, y, params, fold_list, ncores, seed) {
  stopifnot(is.list(params), "k" %in% names(params))
  # initialize a dataframe to store the results
  results_df <- data.table::data.table(
    "fold" = character(0),
    "metric" = numeric(0)
  )
  # we do not need test here as it is defined explicitly below
  params[["test"]] <- NULL
  # loop over the folds
  for (fold in names(fold_list)) {
    # get row-ids of the current fold
    train_idx <- fold_list[[fold]]
    # create learner-arguments
    args <- kdry::list.append(
      list(
        x = kdry::mlh_subset(x, train_idx),
        test = kdry::mlh_subset(x, -train_idx),
        y = kdry::mlh_subset(y, train_idx),
        use.all = FALSE,
        ncores = ncores,
        seed = seed
      ),
      params
    )
    set.seed(seed)
    cvfit <- do.call(knn_fit, args)
    # optimize error rate
    FUN <- metric("ce") # nolint
    err <- FUN(predictions = knn_predict(
      model = cvfit,
      newdata = kdry::mlh_subset(x, -train_idx),
      ncores = ncores,
      type = "response"
      ),
      ground_truth = kdry::mlh_subset(y, -train_idx)
    )
    results_df <- data.table::rbindlist(
      l = list(results_df, list("fold" = fold, "validation_metric" = err)),
      fill = TRUE
    )
  }
  res <- list("metric_optim_mean" = mean(results_df$validation_metric))
  return(res)
}

The bayesian_scoring_function Method

This function can be thought of as a “gatekeeper” that takes a new suggested hyperparameter configuration from the Bayesian process and forwards this configuration further on to a call of the cross_validation method (see above) in order to evaluate this specific setting. However, some peculiarities must be considered in this regard:

  1. The functions needs to take the hyperparameters that should be optimized as function arguments (I generally use the ellipsis (...), however, the hyperparameters can also be defined as arguments explicitly).

  2. When using the strategy = "bayesian", the package is configured in a way that the Bayesian process is parallelized, hence parallel threads evaluate different hyperparameter settings simultaneously (see ParBayesianOptimization's Readme for more details). Therefore, the call to the cross_validation method must explicitly specify ncores = 1L in order to no get in trouble with requesting more resources than available, when using the strategy = "bayesian".

  3. The value returned from the Bayesian scoring function must be a named list that contains the optimization metric as the item Score. As described above, the returned value from cross_validation is already a named list that contains the optimization metric with the item metric_optim_mean. As this item is required later on internally for the mlexperiments package , the value value of this item is just copied and saved under the new name “Score” to address the requirements of ParBayesianOptimization. Note: please notice that mlexperiments already takes care of the direction of the optimization metric, which is handled depending on the learner’s initialization argument metric_optimization_higher_better, so no changes should be made here to ensure a correct functioning.

knn_bsF <- function(...) { # nolint
  params <- list(...)
  # call to knn_optimization here with ncores = 1, since the Bayesian search
  # is parallelized already / "FUN is fitted n times in m threads"
  set.seed(seed)#, kind = "L'Ecuyer-CMRG")
  bayes_opt_knn <- knn_optimization(
    x = x,
    y = y,
    params = params,
    fold_list = method_helper$fold_list,
    ncores = 1L, # important, as bayesian search is already parallelized
    seed = seed
  )
  ret <- kdry::list.append(
    list("Score" = bayes_opt_knn$metric_optim_mean),
    bayes_opt_knn
  )
  return(ret)
}

More details on the package ParBayesianOptimization and on how to define the Bayesian scoring function can be found in its package vignette.

For the parallelization of the Bayesian process, all required functions must be exported to the cluster. To facilitate this, a simple wrapper function can be created that returns a character vector of all custom functions that are called from within the Bayesian scoring function. The following function shows the objects that need to be exported for a correct functioning of the LearnerKnn:

# define the objects / functions that need to be exported to each cluster
# for parallelizing the Bayesian optimization.
knn_ce <- function() {
  c("knn_optimization", "knn_fit", "knn_predict", "metric", ".format_xy")
}

Finally, Create an R6 Class for the Learner

Finally, all of these created functions need to be integrated into a learner object. This is basically done by overwriting the placeholders in an R6 learner that inherits from mlexperiments::MLLearnerBase.

The placeholders are:

Name Type Description
private$fun_fit function A function to fit a model of the respective algorithm. The function must return the fitted model.
private$fun_predict function A function to predict the outcome in new data. The returned value of the predict method should be a vector with the predictions.
private$fun_optim_cv function A function to perform a k-fold cross-validation for one hyperparameter setting. The function must return a named list with at least one item called metric_optim_mean, which contains the cross validated error metric.
private$fun_bayesian_scoring_function function A function that is defined according to the requirements of the ParBayesianOptimization R package. It must return a named list that contains the optimization metric as the item Score.
self$environment field The environment, where to search for the objects that need to be exported to a parallel cluster (required for Bayesian optimization). When the R6 learner is part of an R package, you can write the name of the R package here. Otherwise, -1L (the global environment) might be suitable as long as all objects that are defined in the field cluster_export are available from the global environment.
self$cluster_export field A character vector with the names of objects that need to be exported to each node of a parallel cluster when performing a Bayesian optimization.

These assignments should be done in the initialize() function. The following code example shows the assignment of the previously created functions to the respective functions and fields of the newly created R6 class LearnerKnn:

LearnerKnn <- R6::R6Class( # nolint
  classname = "LearnerKnn",
  inherit = mlexperiments::MLLearnerBase,
  public = list(
    initialize = function() {
      if (!requireNamespace("class", quietly = TRUE)) {
        stop(
          paste0(
            "Package \"class\" must be installed to use ",
            "'learner = \"LearnerKnn\"'."
          ),
          call. = FALSE
        )
      }
      super$initialize(
        metric_optimization_higher_better = FALSE # classification error
      )

      private$fun_fit <- knn_fit
      private$fun_predict <- knn_predict
      private$fun_optim_cv <- knn_optimization
      private$fun_bayesian_scoring_function <- knn_bsF

      self$environment <- "mlexperiments"
      self$cluster_export <- knn_ce()
    }
  )
)

Please note that metric_optimization_higher_better is defaulted to FALSE here when initializing the super-class. This is because of choosing the error rate as the optimization metric (FUN <- metric("ce")) when defining the cross_validation-function above.

Examples

Now, the learner is put together and ready to be used with mlexperiments:

Preparations

First of all, load the data and transform it into a matrix, and define the training data and the target variable.

library(mlexperiments)
library(mlbench)

data("DNA")
dataset <- DNA |>
  data.table::as.data.table() |>
  na.omit()

seed <- 123
feature_cols <- colnames(dataset)[1:180]

train_x <- model.matrix(
  ~ -1 + .,
  dataset[, .SD, .SDcols = feature_cols]
)
train_y <- dataset[, get("Class")]

ncores <- ifelse(
  test = parallel::detectCores() > 4,
  yes = 4L,
  no = ifelse(
    test = parallel::detectCores() < 2L,
    yes = 1L,
    no = parallel::detectCores()
  )
)
if (isTRUE(as.logical(Sys.getenv("_R_CHECK_LIMIT_CORES_")))) {
  # on cran
  ncores <- 2L
}

Hyperparameter Tuning

Bayesian Tuning

For the Bayesian hyperparameter optimization, it is required to define a grid with some hyperparameter combinations that is used for initializing the Bayesian process. Furthermore, the borders (allowed extreme values) of the hyperparameters that are actually optimized need to be defined in a list. Finally, further arguments that are passed to the function ParBayesianOptimization::bayesOpt() can be defined as well.

param_list_knn <- expand.grid(
  k = seq(4, 68, 8),
  l = 0,
  test = parse(text = "fold_test$x")
)

knn_bounds <- list(k = c(2L, 80L))

optim_args <- list(
  iters.n = ncores,
  kappa = 3.5,
  acq = "ucb"
)

Here, another peculiarity of class::knn() is visible: when fitting a model, one needs to specify the argument test in order to specify a matrix of test set cases. In order to have the correct test set cases selected throughout the cross-validation, one needs to specify argument as an expression, which is then evaluated before passing the arguments on to the fit-function.

Generally speaking, this is a feature implemented in mlexperiments: when specifying an expression as a learner argument (either via the R6 classes’ fields learner_args or parameter_grid), this expression is evaluated before passing the argument list on the fitting functions.

In order to execute the parameter tuning, the created objects need to be assigned to the corresponding fields of the R6 class mlexperiments::MLTuneParameters:

knn_tune_bayesian <- mlexperiments::MLTuneParameters$new(
  learner = LearnerKnn$new(),
  strategy = "bayesian",
  ncores = ncores,
  seed = seed
)

knn_tune_bayesian$parameter_bounds <- knn_bounds
knn_tune_bayesian$parameter_grid <- param_list_knn
knn_tune_bayesian$split_type <- "stratified"
knn_tune_bayesian$optim_args <- optim_args

# set data
knn_tune_bayesian$set_data(
  x = train_x,
  y = train_y
)

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

head(results)
#>    Epoch setting_id  k gpUtility acqOptimum inBounds Elapsed      Score metric_optim_mean errorMessage l
#> 1:     0          1  4        NA      FALSE     TRUE   2.153 -0.2247332         0.2247332           NA 0
#> 2:     0          2 12        NA      FALSE     TRUE   2.274 -0.1600753         0.1600753           NA 0
#> 3:     0          3 20        NA      FALSE     TRUE   2.006 -0.1381042         0.1381042           NA 0
#> 4:     0          4 28        NA      FALSE     TRUE   2.329 -0.1403013         0.1403013           NA 0
#> 5:     0          5 36        NA      FALSE     TRUE   2.109 -0.1315129         0.1315129           NA 0
#> 6:     0          6 44        NA      FALSE     TRUE   2.166 -0.1258632         0.1258632           NA 0

To carry out the hyperparameter optimization with a grid search, only the parameter_grid is required:

knn_tune_grid <- mlexperiments::MLTuneParameters$new(
  learner = LearnerKnn$new(),
  strategy = "grid",
  ncores = ncores,
  seed = seed
)

knn_tune_grid$parameter_grid <- param_list_knn
knn_tune_grid$split_type <- "stratified"

# set data
knn_tune_grid$set_data(
  x = train_x,
  y = train_y
)

results <- knn_tune_grid$execute(k = 3)
#> 
#> Parameter settings [=====================>---------------------------------------------------------------------------] 2/9 ( 22%)
#> Parameter settings [===============================>-----------------------------------------------------------------] 3/9 ( 33%)
#> Parameter settings [==========================================>------------------------------------------------------] 4/9 ( 44%)
#> Parameter settings [=====================================================>-------------------------------------------] 5/9 ( 56%)
#> Parameter settings [================================================================>--------------------------------] 6/9 ( 67%)
#> Parameter settings [==========================================================================>----------------------] 7/9 ( 78%)
#> Parameter settings [=====================================================================================>-----------] 8/9 ( 89%)
#> Parameter settings [=================================================================================================] 9/9 (100%)                                                                                                                                  

head(results)
#>    setting_id metric_optim_mean  k l
#> 1:          1         0.2187696  4 0
#> 2:          2         0.1597615 12 0
#> 3:          3         0.1349655 20 0
#> 4:          4         0.1406152 28 0
#> 5:          5         0.1318267 36 0
#> 6:          6         0.1258632 44 0

Cross-Validation

For the cross-validation experiments (mlexperiments::MLCrossValidation, and mlexperiments::MLNestedCV), a named list with the in-sample row indices of the folds is required.

fold_list <- splitTools::create_folds(
  y = train_y,
  k = 3,
  type = "stratified",
  seed = seed
)
str(fold_list)
#> List of 3
#>  $ Fold1: int [1:2124] 1 2 3 4 5 7 9 10 11 12 ...
#>  $ Fold2: int [1:2124] 1 2 3 6 8 9 11 13 16 17 ...
#>  $ Fold3: int [1:2124] 4 5 6 7 8 10 12 14 15 16 ...

Furthermore, a specific hyperparameter setting that should be validated with the cross-validation needs to be selected:

knn_cv <- mlexperiments::MLCrossValidation$new(
  learner = LearnerKnn$new(),
  fold_list = fold_list,
  seed = seed
)

best_grid_result <- knn_tune_grid$results$best.setting
best_grid_result
#> $setting_id
#> [1] 9
#> 
#> $k
#> [1] 68
#> 
#> $l
#> [1] 0
#> 
#> $test
#> expression(fold_test$x)

knn_cv$learner_args <- best_grid_result[-1]

knn_cv$predict_args <- list(type = "response")
knn_cv$performance_metric <- metric("bacc")
knn_cv$return_models <- TRUE

# set data
knn_cv$set_data(
  x = train_x,
  y = train_y
)

results <- knn_cv$execute()
#> 
#> CV fold: Fold1
#> 
#> CV fold: Fold2
#> CV progress [====================================================================>-----------------------------------] 2/3 ( 67%)
#> 
#> CV fold: Fold3
#> CV progress [========================================================================================================] 3/3 (100%)
#>                                                                                                                                   

head(results)
#>     fold performance  k l
#> 1: Fold1   0.8912781 68 0
#> 2: Fold2   0.8832388 68 0
#> 3: Fold3   0.8657147 68 0

Nested Cross-Validation

Last but not least, the hyperparameter optimization and validation can be combined in a nested cross-validation. In each fold of the so-called “outer” cross-validation loop, the hyperparameters are optimized on the in-sample observations with one of the two strategies: Bayesian optimization or grid search. Both of these strategies are implemented again with a “nested” (“inner”) cross-validation. The best hyperparameter setting as identified by the inner cross-validation is then used to fit a model with all in-sample observations of the outer cross-validation loop and finally validate it on the respective out-sample observations.

The experiment classes must be parameterized as described above.

Inner Bayesian Optimization

knn_cv_nested_bayesian <- mlexperiments::MLNestedCV$new(
  learner = LearnerKnn$new(),
  strategy = "bayesian",
  fold_list = fold_list,
  k_tuning = 3L,
  ncores = ncores,
  seed = seed
)

knn_cv_nested_bayesian$parameter_grid <- param_list_knn
knn_cv_nested_bayesian$parameter_bounds <- knn_bounds
knn_cv_nested_bayesian$split_type <- "stratified"
knn_cv_nested_bayesian$optim_args <- optim_args

knn_cv_nested_bayesian$predict_args <- list(type = "response")
knn_cv_nested_bayesian$performance_metric <- metric("bacc")

# set data
knn_cv_nested_bayesian$set_data(
  x = train_x,
  y = train_y
)

results <- knn_cv_nested_bayesian$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(results)
#>     fold performance  k l
#> 1: Fold1   0.8912781 68 0
#> 2: Fold2   0.8832388 68 0
#> 3: Fold3   0.8657147 68 0
knn_cv_nested_grid <- mlexperiments::MLNestedCV$new(
  learner = LearnerKnn$new(),
  strategy = "grid",
  fold_list = fold_list,
  k_tuning = 3L,
  ncores = ncores,
  seed = seed
)

knn_cv_nested_grid$parameter_grid <- param_list_knn
knn_cv_nested_grid$split_type <- "stratified"

knn_cv_nested_grid$predict_args <- list(type = "response")
knn_cv_nested_grid$performance_metric <- metric("bacc")

# set data
knn_cv_nested_grid$set_data(
  x = train_x,
  y = train_y
)

results <- knn_cv_nested_grid$execute()
#> 
#> CV fold: Fold1
#> 
#> Parameter settings [=====================>---------------------------------------------------------------------------] 2/9 ( 22%)
#> Parameter settings [===============================>-----------------------------------------------------------------] 3/9 ( 33%)
#> Parameter settings [==========================================>------------------------------------------------------] 4/9 ( 44%)
#> Parameter settings [=====================================================>-------------------------------------------] 5/9 ( 56%)
#> Parameter settings [================================================================>--------------------------------] 6/9 ( 67%)
#> Parameter settings [==========================================================================>----------------------] 7/9 ( 78%)
#> Parameter settings [=====================================================================================>-----------] 8/9 ( 89%)
#> Parameter settings [=================================================================================================] 9/9 (100%)                                                                                                                                  
#> CV fold: Fold2
#> CV progress [====================================================================>-----------------------------------] 2/3 ( 67%)
#> 
#> Parameter settings [=====================>---------------------------------------------------------------------------] 2/9 ( 22%)
#> Parameter settings [===============================>-----------------------------------------------------------------] 3/9 ( 33%)
#> Parameter settings [==========================================>------------------------------------------------------] 4/9 ( 44%)
#> Parameter settings [=====================================================>-------------------------------------------] 5/9 ( 56%)
#> Parameter settings [================================================================>--------------------------------] 6/9 ( 67%)
#> Parameter settings [==========================================================================>----------------------] 7/9 ( 78%)
#> Parameter settings [=====================================================================================>-----------] 8/9 ( 89%)
#> Parameter settings [=================================================================================================] 9/9 (100%)                                                                                                                                  
#> CV fold: Fold3
#> CV progress [========================================================================================================] 3/3 (100%)
#>                                                                                                                                   
#> Parameter settings [=====================>---------------------------------------------------------------------------] 2/9 ( 22%)
#> Parameter settings [===============================>-----------------------------------------------------------------] 3/9 ( 33%)
#> Parameter settings [==========================================>------------------------------------------------------] 4/9 ( 44%)
#> Parameter settings [=====================================================>-------------------------------------------] 5/9 ( 56%)
#> Parameter settings [================================================================>--------------------------------] 6/9 ( 67%)
#> Parameter settings [==========================================================================>----------------------] 7/9 ( 78%)
#> Parameter settings [=====================================================================================>-----------] 8/9 ( 89%)
#> Parameter settings [=================================================================================================] 9/9 (100%)                                                                                                                                  

head(results)
#>     fold performance  k l
#> 1: Fold1   0.8959736 52 0
#> 2: Fold2   0.8832388 68 0
#> 3: Fold3   0.8657147 68 0