Definition

In machine learning models, feature importance refers to methods of assigning scores to input features. These scores represent how signficant each feature is in influencing the model's predictions.

Feature importance allows you to:

• Improve understanding - Recognise key factors driving a model's decisions.
• Reduce dimensionality - Remove irrelevant or redundant features to streamline models.
• Debug models - Identify features that may be negatively impacting performance.

Common Techniques

Impurity-based feature importance (Decision Trees, Random Forests)

Calculates the reduction in node impurity (e.g. Gini Impurity, Entropy) brought about by each feature split across all trees in an ensemble. Features that lead to larger impurity reductions are deemed more important.

It is built-in to decision tree-based methods, and is computationally efficient.

It can be biased towards features with more categories or continuous features.

Permutation importance

For each feature, its values are randomly shuffled, and the decrease in model performance (e.g. accuracy decline) is measured. Features leading to a larger drop in performance when shuffled are considered more important.

It is model-agnostic, and also considers feature interactions.

It is computationally more expensive than impurity-based importance.

Sensitivity analysis

Measures how small changes in individual features impact the model's output. Largest changes in output indicate higher feature importance.

It provides insight into the direction of the feature's impact (positive or negative).

Computationally intensive, especially with large number of features.

SHAP (SHapley Additive exPlanations)

Model-agnostic method based on game theory. Uses Shapley values to determine the fair distribution of attribution of a prediction to each feature.

Considers feature interactions and can handle feature dependence, provides global and local explanations.

It is computationally expensive.

Technique Selection

• Model Type - If you're using decision tress or tree ensembles, impurity-base importance is a natural choice. Otherwise, SHAP or permutation importance are more versatile.
• Computation Cost - Permutation importance and SHAP can be slower, especially for a large dataset. If speed is crucial, impurity-based or sensitivity analysis might be better.
• Need for Global vs. Local Explanations - SHAP is excellent for both understanding features' overall importance and for explaining individual predictions.

Libraries for Implementation

• scikit-learn (Python) - Provides impurity-based importance for tree models, permutation importance, and limited partial dependence plots for sensitivity.
• SHAP (Python) - A powerful library specifically designed for calculating SHAP values.