Extract outlier configurations

Extracts hyperparameter configurations identified as outliers during subspace fitting. Outliers are configurations that violated the subspace constraints and were excluded via slack variables during regularized optimization.

Usage

outliers(object, ...)

Arguments

object

A trained LearnerSubspace object (or subclass) with fitted subspace parameters

...

Additional arguments (currently unused)

Value

A data.table containing the outlier configurations with all columns from the original task data. Returns an empty data.table if no outliers were identified. A message is printed when no outliers exist.

Details

When Outliers Exist:

Outliers are only identified when training with lambda > 0. The regularization parameter allows the optimization to exclude some configurations from the fitted subspace by introducing slack variables.

For a configuration to be considered an outlier, its slack variable must exceed the threshold of \(10^{-5}\).

When lambda = NULL, all configurations are forced to fit within the subspace (hard constraints), so no outliers exist.

Categorical Hyperparameters:

When the task includes categorical hyperparameters, outliers are identified separately for each categorical level and combined in the returned data.table.

Interpretation:

Outliers typically represent:

• Configurations in sparse regions of hyperparameter space

• Anomalous configurations with unusual performance

• Configurations that don't fit the dominant subspace pattern

The number and characteristics of outliers can guide decisions about:

• Adjusting the lambda parameter

• Investigating unusual configurations

• Understanding the geometry of high-performing regions

See also

LearnerSubspace for the base learner class. LearnerSubspaceBox for axis-aligned hyperrectangles. LearnerSubspacePolygon for oriented hyperrectangles. LearnerSubspaceEllipsoid for ellipsoids.

Examples

if (FALSE) { # \dontrun{
# Train with regularization (allows outliers)
task <- TaskSubspace$new(data, target_measure = "auc",
                         hps = c("learning_rate", "max_depth"))
learner <- LearnerSubspaceBox$new(task)
learner$train(q_val = 0.9, lambda = 0.1)

# Extract outliers
outlier_configs <- outliers(learner)
print(nrow(outlier_configs))  # Number of outliers
print(outlier_configs)  # View outlier configurations

# Check outlier information from result
print(learner$result$n_violations)  # Total number of outliers
print(learner$result$outliers)  # Outlier indices

# Train without regularization (no outliers)
learner$train(q_val = 0.9, lambda = NULL)
outlier_configs <- outliers(learner)  # Returns empty data.table

# With categorical hyperparameters
task <- TaskSubspace$new(data, target_measure = "auc",
                         hps = c("learning_rate", "max_depth"),
                         cat_hps = "optimizer")
learner <- LearnerSubspaceBox$new(task)
learner$train(q_val = 0.9, lambda = 0.1)
outlier_configs <- outliers(learner)  # Combined across all levels
} # }