mlr_filters_ensemble: Filter Ensemble

In mlr3pipelines: Preprocessing Operators and Pipelines for 'mlr3'

Filter Ensemble

Description

FilterEnsemble aggregates several Filters by averaging their scores (or ranks) with user-defined weights. Each wrapped filter is evaluated on the supplied task, and the resulting feature scores are combined feature-wise by a convex combination determined through the weights parameter. This allows leveraging complementary inductive biases of multiple filters without committing to a single criterion. The concept was introduced by Binder et al. (2020). This implementation follows the idea but leaves the exact choice of weights to the user.

Format

R6Class object inheriting from Filter.

Construction

FilterEnsemble$new(filters)

• filters :: list of Filter
  Filters that are evaluated and aggregated. Each filter must be cloneable and support the task type and feature types of the ensemble. The ensemble identifier defaults to the wrapped filter ids concatenated by ".".

Parameters

• weights :: numeric()
  Required non-negative weights, one for each wrapped filter, with at least one strictly positive value. Values are used as given when calculating the weighted mean. If named, names must match the wrapped filter ids.

• rank_transform :: logical(1)
  If TRUE, ranks of individual filter scores are used instead of the raw scores before averaging. Initialized to FALSE.

Parameters of wrapped filters are available via ⁠$param_set⁠ and can be referenced using the wrapped filter id followed by ".", e.g. "variance.na.rm".

Fields

• ⁠$wrapped⁠ :: named list of Filter
  Read-only access to the wrapped filters.

Methods

• get_weights_search_space(weights_param_name = "weights", normalize_weights = "uniform", prefix = "w")
  (character(1), character(1), character(1)) -> ParamSet
  Construct a ParamSet describing a weight search space.

• get_weights_tunetoken(normalize_weights = "uniform")
  (character(1)) -> TuneToken
  Shortcut returning a TuneToken for tuning the weights.

• set_weights_to_tune(normalize_weights = "uniform")
  (character(1)) -> self
  Convenience wrapper that stores the TuneToken returned by get_weights_tunetoken() in ⁠$param_set$values$weights⁠.

Internals

All wrapped filters are called with nfeat equal to the number of features to ensure that complete score vectors are available for aggregation. Scores are combined per feature by computing the weighted (optionally rank-based) mean.

References

Binder M, Moosbauer J, Thomas J, Bischl B (2020). “Multi-objective hyperparameter tuning and feature selection using filter ensembles.” In Proceedings of the 2020 Genetic and Evolutionary Computation Conference, 471–479. \Sexpr[results=rd]{tools:::Rd_expr_doi("10.1145/3377930.3389815")}.

Examples

library("mlr3")
library("mlr3filters")

task = tsk("sonar")

flt = mlr_filters$get("ensemble",
  filters = list(FilterVariance$new(), FilterAUC$new()))
flt$param_set$values$weights = c(variance = 0.5, auc = 0.5)
flt$calculate(task)
head(as.data.table(flt))

mlr3pipelines documentation built on Nov. 7, 2025, 9:06 a.m.