Stability-Based Ensemble Selection

Builds an ensemble of models using stability-based feature selection. The approach: 1. Run N bootstrap resamples with feature selection 2. Compute feature selection frequencies 3. Create nested overlapping subsets based on stability tiers 4. Train K base models on different stability tiers 5. Combine predictions via stability-weighted voting

This approach addresses the "reproducibility crisis" in biomarker discovery by prioritizing stable, reproducible feature sets over single-run selections.

References

Meinshausen & Buhlmann (2010). Stability Selection. JRSS-B. Nogueira et al. (2018). On the Stability of Feature Selection Algorithms.

Public fields

n_bootstrap

Number of bootstrap resamples

filter

Feature selection filter method

n_features

Number of features to select per bootstrap

base_learner

Base learner for ensemble members

threshold_mode

"adaptive" (target set size) or "fixed" (percentage)

target_set_size

Target number of stable features (for adaptive mode)

fixed_thresholds

Thresholds for fixed mode (creates nested subsets)

aggregation

"weighted" (stability-weighted) or "average" (simple)

verbose

Print progress messages

frequencies

Feature selection frequencies from last fit

stable_features

List of stable feature sets by tier

tier_weights

Weights for each tier in ensemble

fitted_models

List of fitted models per tier

Methods

Public methods

• StabilityEnsemble$new()

• StabilityEnsemble$compute_frequencies()

• StabilityEnsemble$identify_stable_subsets()

• StabilityEnsemble$create_tier_learner()

• StabilityEnsemble$fit()

• StabilityEnsemble$predict()

• StabilityEnsemble$get_feature_importance()

• StabilityEnsemble$get_summary()

• StabilityEnsemble$print()

• StabilityEnsemble$clone()

---

Method new()

Create a new StabilityEnsemble

Usage

StabilityEnsemble$new(
  n_bootstrap = 100,
  filter = "anova",
  n_features = 50,
  base_learner = "glmnet",
  threshold_mode = "adaptive",
  target_set_size = c(10, 50),
  fixed_thresholds = c(0.9, 0.7, 0.5),
  aggregation = "weighted",
  verbose = TRUE
)

Arguments

n_bootstrap

Number of bootstrap resamples (default: 100)

filter

Feature selection filter: "anova", "mrmr", "importance"

n_features

Features to select per bootstrap (default: 50)

base_learner

Learner for ensemble: "glmnet", "ranger", "xgboost"

threshold_mode

"adaptive" or "fixed"

target_set_size

Target stable set size range for adaptive mode

fixed_thresholds

Thresholds for fixed mode

aggregation

"weighted" or "average"

verbose

Print progress

Returns

A new StabilityEnsemble object

---

Method compute_frequencies()

Compute feature selection frequencies from bootstrap resamples

Usage

StabilityEnsemble$compute_frequencies(task, seed = NULL)

Arguments

task

An mlr3 Task

seed

Random seed for reproducibility

Returns

Named numeric vector of selection frequencies

---

Method identify_stable_subsets()

Identify stable feature subsets based on frequency thresholds

Usage

StabilityEnsemble$identify_stable_subsets()

Returns

List of feature vectors for each stability tier

---

Method create_tier_learner()

Create mlr3 learner for a specific feature subset

Usage

StabilityEnsemble$create_tier_learner(features)

Arguments

features

Character vector of feature names

Returns

An mlr3 Learner

---

Method fit()

Fit the stability ensemble on a task

Usage

StabilityEnsemble$fit(task, seed = NULL)

Arguments

task

An mlr3 Task

seed

Random seed

Returns

The fitted StabilityEnsemble (invisibly)

---

Method predict()

Predict on new data using the ensemble

Usage

StabilityEnsemble$predict(task)

Arguments

task

An mlr3 Task (or data.frame with same features)

Returns

A data.table with predicted probabilities

---

Method get_feature_importance()

Get feature importance based on selection frequencies

Usage

StabilityEnsemble$get_feature_importance(top_n = NULL)

Arguments

top_n

Number of top features to return (default: all)

Returns

Named vector of feature frequencies, sorted descending

---

Method get_summary()

Get summary of the ensemble configuration and results

Usage

StabilityEnsemble$get_summary()

Returns

A list with ensemble summary

---

Method print()

Print summary

Usage

StabilityEnsemble$print()

---

Method clone()

The objects of this class are cloneable with this method.

Usage

StabilityEnsemble$clone(deep = FALSE)

Arguments

deep

Whether to make a deep clone.