Leave-One-Covariate-In (LOCI)

Calculates Leave-One-Covariate-In (LOCI) scores. This method is primariliy implemented for completeness' sake, but in general there are more informative measures for the marginal association between a feature and the target.

Details

LOCI measures feature importance by training models with only each individual feature and comparing their performance to a featureless baseline model (optimal constant prediction). The importance is calculated as (featureless_model_loss - single_feature_loss). Positive values indicate the feature performs better than the baseline, negative values indicate worse performance.

Super classes

xplainfi::FeatureImportanceMethod -> xplainfi::LeaveOutIn -> LOCI

Methods

Public methods

• LOCI$new()

• LOCI$clone()

Inherited methods

• xplainfi::FeatureImportanceMethod$importance()
• xplainfi::FeatureImportanceMethod$print()
• xplainfi::FeatureImportanceMethod$reset()
• xplainfi::LeaveOutIn$compute()

---

Method new()

Creates a new instance of this R6 class.

Usage

LOCI$new(
  task,
  learner,
  measure,
  resampling = NULL,
  features = NULL,
  iters_refit = 1L,
  obs_loss = FALSE
)

Arguments

task

(mlr3::Task) Task to compute importance for.

learner

(mlr3::Learner) Learner to use for prediction.

measure

(mlr3::Measure) Measure to use for scoring.

resampling

(mlr3::Resampling) Resampling strategy. Defaults to holdout.

features

(character()) Features to compute importance for. Defaults to all features.

iters_refit

(integer(1)) Number of refit iterations per resampling iteration.

obs_loss

(logical(1)) Whether to use observation-wise loss calculation (analogous to LOCO) when supported by the measure. If FALSE (default), uses aggregated scores. When TRUE, uses the measure's aggregation function or mean as fallback.

---

Method clone()

The objects of this class are cloneable with this method.

Usage

LOCI$clone(deep = FALSE)

Arguments

deep

Whether to make a deep clone.

Examples

library(mlr3)
library(mlr3learners)
task = tgen("friedman1")$generate(n = 200)

# Standard LOCI with aggregated scores
loci = LOCI$new(
  task = task,
  learner = lrn("regr.ranger", num.trees = 50),
  measure = msr("regr.mse")
)
#> ℹ No <Resampling> provided
#> Using `resampling = rsmp("holdout")` with default `ratio = 0.67`.
loci$compute()

# Using observation-wise losses with measure's aggregation function
loci_obsloss = LOCI$new(
  task = task,
  learner = lrn("regr.ranger", num.trees = 50),
  measure = msr("regr.mae"), # uses MAE's aggregation function (mean) internally
  obs_loss = TRUE
)
#> ℹ No <Resampling> provided
#> Using `resampling = rsmp("holdout")` with default `ratio = 0.67`.
loci_obsloss$compute()
loci_obsloss$obs_losses
#>      row_ids      feature iteration iter_refit    truth response_ref
#>        <int>       <char>     <int>      <int>    <num>        <num>
#>   1:       4   important1         1          1 15.36222     14.04979
#>   2:       4   important2         1          1 15.36222     14.04979
#>   3:       4   important3         1          1 15.36222     14.04979
#>   4:       4   important4         1          1 15.36222     14.04979
#>   5:       4   important5         1          1 15.36222     14.04979
#>  ---                                                                
#> 666:     190 unimportant1         1          1 11.60406     14.04979
#> 667:     190 unimportant2         1          1 11.60406     14.04979
#> 668:     190 unimportant3         1          1 11.60406     14.04979
#> 669:     190 unimportant4         1          1 11.60406     14.04979
#> 670:     190 unimportant5         1          1 11.60406     14.04979
#>      response_feature loss_ref loss_feature   obs_diff
#>                 <num>    <num>        <num>      <num>
#>   1:         15.42979 1.312421   0.06757879  1.2448426
#>   2:         12.11727 1.312421   3.24494453 -1.9325232
#>   3:         20.14413 1.312421   4.78191717 -3.4694958
#>   4:         11.69276 1.312421   3.66945606 -2.3570347
#>   5:         13.48251 1.312421   1.87971032 -0.5672890
#>  ---                                                  
#> 666:         13.96568 2.445737   2.36161915  0.0841178
#> 667:         18.78093 2.445737   7.17687679 -4.7311398
#> 668:         13.16344 2.445737   1.55938273  0.8863542
#> 669:         15.96397 2.445737   4.35991736 -1.9141804
#> 670:         12.86600 2.445737   1.26193892  1.1837980

# LOCI with median aggregation (analogous to original LOCO)
mae_median = msr("regr.mae")
mae_median$aggregator = median
loci_median = LOCI$new(
  task = task,
  learner = lrn("regr.ranger", num.trees = 50),
  measure = mae_median,
  obs_loss = TRUE
)
#> ℹ No <Resampling> provided
#> Using `resampling = rsmp("holdout")` with default `ratio = 0.67`.
loci_median$compute()