Package index • xplainfi

Importance Measures

Base class

FeatureImportanceMethod: Feature Importance Method Class

Perturbation-Based Importance Measures

Methods which perturb features of interest either marginally (PFI) or conditionally on all (CFI) or a subset of remaining features (RFI)

PerturbationImportance: Perturbation Feature Importance Base Class
PFI: Permutation Feature Importance
CFI: Conditional Feature Importance
RFI: Relative Feature Importance

Model Refitting Measures

Methods which refit models with one (or more) features omitted (LOCO) or included (LOCI).

WVIM: Williamson's Variable Importance Measure (WVIM)
LOCO: Leave-One-Covariate-Out (LOCO)

Shapley-Based Approaches

Shapley Additive Global Importance (SAGE) in marginal and conditional variants

SAGE: Shapley Additive Global Importance (SAGE) Base Class
MarginalSAGE: Marginal SAGE
ConditionalSAGE: Conditional SAGE

Sampling Infrastructure

Base class and three families of samplers (Marginal, Conditional, Knockoff)

Base classes

Abstract base classes for the three sampler families. FeatureSampler is the top-level base class, MarginalSampler is for marginal sampling methods, and ConditionalSampler is for conditional sampling methods.

FeatureSampler: Feature Sampler Class
MarginalSampler: Marginal Sampler Base Class
ConditionalSampler: Conditional Feature Sampler

Marginal sampling (no conditioning)

Samplers that draw from the marginal distribution P(X_S) without conditioning on other features.

MarginalPermutationSampler: Marginal Permutation Sampler
MarginalReferenceSampler: Marginal Reference Sampler

Conditional sampling (with conditioning)

Samplers that draw from the conditional distribution P(X_S | X_C) where X_C is an arbitrary conditioning set. The conditioning set can be specified via the conditioning_set parameter.

ConditionalARFSampler: ARF-based Conditional Sampler
ConditionalGaussianSampler: Gaussian Conditional Sampler
ConditionalKNNSampler: k-Nearest Neighbors Conditional Sampler
ConditionalCtreeSampler: (experimental) Conditional Inference Tree Conditional Sampler

Knockoff sampling

Knockoffs satisfy certain theoretical properties that exceed those of the conditional samplers. They generate one (or more) knockoff matrix on construction, and sampling is always performed in reference to these. Unlike other samplers, they do not allow sampling from new data with $sample_newdata(), as that would require re-creating a knockoff matrix on the fly.

KnockoffSampler: Knockoff-based Conditional Sampler
KnockoffGaussianSampler: Gaussian Knockoff Conditional Sampler
KnockoffSequentialSampler: Gaussian Knockoff Conditional Sampler

Utilities

wvim_design_matrix(): Create Feature Selection Design Matrix
check_groups(): Check group specification
`%||%`: Default value for NULL

Data simulation

sim_dgp_ewald(): Simulate data as in Ewald et al. (2024)
sim_dgp_correlated() sim_dgp_mediated() sim_dgp_confounded() sim_dgp_interactions() sim_dgp_independent(): Simulation DGPs for Feature Importance Method Comparison