Permutation feature importance

As explained in Ensemble Methods for Machine Learning

Permutation feature importance estimates a feature's contribution by measuring the change in model performance after that feature's values are randomly permuted.

Permutation feature importance estimates a feature’s contribution by measuring how much model performance worsens after that feature’s values are randomly shuffled. Shuffling breaks the relationship between the feature and predictions without requiring access to model internals, making the method model-agnostic. section 9.3.1

Compute the score as a before-and-after comparison:

• Evaluate the trained model on a hold-out test set. section 9.3.1
• Randomly permute one feature’s values across the test examples. section 9.3.1
• Evaluate the model again and measure the increase in test error, repeating the process for every feature. section 9.3.1
• Interpret larger performance degradation as evidence that the feature contributes more to predictions; a small change suggests weaker contribution. section 9.3.1

XGBoost with scikit-learn

Apply permutation importance to an XGBoost model with scikit-learn’s permutation_importance function, using a hold-out test set, balanced accuracy as the scoring metric, and 30 repeats. Different metrics can produce different importance results, such as balanced accuracy versus recall. section 9.3.1

Figure 9.7 The procedure for computing permutation feature importance illustrated for the third feature. This procedure is repeated for all features. Permutation feature importance uses only inputs and outputs to estimate feature importance and doesn’t use model internals (making this a model-agnostic approach).

Follow the visual procedure for permuting one feature, measuring the performance change, and repeating the calculation across all features. section 9.3.1

How Ensemble Methods for Machine Learning covers it

How deep it goes

Explains the shuffle-and-reevaluate procedure, test-set requirement, metric choice, and repeated computation across features.

Examples

Uses an XGBoost model evaluated with balanced accuracy or recall and 30 permutation repeats.

What it assumes

Assumes a trained predictive model, a hold-out test set, and a task-appropriate performance metric.

What you can do after

Apply it via scikit-learn to rank features by their effect on hold-out predictive performance.

How it connects

Contrasts with decision-tree feature importance and complements partial dependence plots, surrogate models, LIME, and SHAP.

liveBookmanning.com

Read Ensemble Methods for Machine Learning on liveBook Get the book at manning.com

Also explained in

• 

  liveBookmanning.com
  Machine Learning System DesignPermutation feature importance measures a feature's importance by randomly shuffling its values and measuring the resulting change in model performance.

Compare how 2 books explain permutation feature importance →

Related concepts

• importance scorekind of · An importance score is a numerical value indicating an item's relative significance, relevance, or contribution within a specified context.
• out-of-bag erroruses · Out-of-bag error is an estimate of an ensemble’s prediction error computed for each observation using only ensemble members whose bootstrap samples omitted it.
• independent variablemeasures · An independent variable is an input or explanatory attribute whose values are used to model, explain, or predict variation in an outcome.
• test errormeasures · Test error is the frequency or magnitude of a model’s prediction errors on data held out from fitting for evaluation.
• model-agnostic methodkind of · A model-agnostic method analyzes predictions or estimates explanations without relying on a model’s internal structure, implementation, or architecture.
• evaluation datasetuses · An evaluation dataset is a collection of examples or test cases, often with expected or judged outcomes, designated to assess a model or system against defined criteria.

This explainer was generated by Manning's knowledge graph from the text of Ensemble Methods for Machine Learning. The section links point to the parts of the book it draws on.