Research map / Probabilistic and scientific modeling / Molecules and proteins

Conditional Feature Importance with Generative Modeling Using Adversarial Random Forests

AAAI 2025 · Probabilistic and scientific modeling › Molecules and proteins · approach: Sparse concepts and hypothesis testing · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes cARFi, a model-agnostic conditional feature importance measure that samples on-manifold feature values from adversarial random forest (ARF) conditional distributions, with tests for nonzero importance.

Problem

Measuring conditional (and relative) feature importance for a fitted model requires sampling feature values from conditional distributions, which is hard for mixed tabular data; existing methods (SAGE, CS, CPI with knockoffs) are costly, lack flexible conditioning, or lack valid inference.

Approach

Use an ARF, a fast tree-based generative model fitted once on features only, to derive conditional densities for arbitrary conditioning sets without refitting, sample R replicates per instance for stable loss differences, and tune minimum node size to interpolate between marginal and conditional importance; asymptotic normality gives paired t-tests.

Key results

• In the proof of concept, one-sided paired t-tests at alpha=0.05 control type I error near 5% at effect size 0 and reach 100% power at larger effect sizes for most learners.
• In the mixed-data DAG simulation, cARFi with R=20 keeps type I error controlled for unimportant features and shows good power, improving on single-sample CPI knockoffs especially at small sample sizes.
• cARFi is reported to be faster than generating sequential knockoffs, with ARF sampling much faster at R=20 (runtime in supplement).

Measured effects reported in the paper

Surprising findings

• Very deep trees (small minimum leaf size) give conservative tests, while shallow trees inflate type I error, so the tree-depth parameter acts as an in-between marginal-conditional knob.

Limitations

• Asymptotic normality guaranteed only for models with finite VC dimension; SVM results are empirical
• Uncertainty in the ARF conditional distributions and the learned model is ignored in the statistical tests
• Evaluation is limited to simulations and one real dataset; no sensitivity analysis over sample size, dimension, or hyperparameters

Experimental setup

Datasets: simulated Gaussian data simulated mixed categorical/Gaussian DAG data bike-sharing dataset

Benchmarks: simulated linear and nonlinear settings (proof of concept) mixed data DAG simulation bike-sharing dataset

Metrics: type I error rate statistical power feature importance values runtime

Baselines: CPI with sequential knockoffs CPI with Gaussian knockoffs conditional subgroups (CS) SAGE PFI KernelSHAP

Builds on: adversarial random forest conditional predictive impact permutation feature importance model-X knockoffs SAGE

Closest papers by meaning

• Testing Semantic Importance via Betting (NeurIPS 2024 · similarity 0.86)
• Accurate Estimation of Feature Importance Faithfulness for Tree Models (AAAI 2025 · similarity 0.84)
• Feature Importance Metrics in the Presence of Missing Data (ICML 2025 · similarity 0.84)
• Variable Importance in High-Dimensional Settings Requires Grouping (AAAI 2024 · similarity 0.83)
• A Likelihood Based Approach to Distribution Regression Using Conditional Deep Generative Models (ICML 2025 · similarity 0.83)
• Generative Classifiers Avoid Shortcut Solutions (ICLR 2025 · similarity 0.83)