Research map / Probabilistic and scientific modeling / Variational and generative sampling

GEFA: A General Feature Attribution Framework Using Proxy Gradient Estimation

ICML 2025 · Probabilistic and scientific modeling › Variational and generative sampling · approach: Gradient estimation for discrete variables

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 GEFA, a black-box feature attribution framework that estimates gradients in a proxy (Bernoulli presence-probability) space, works for discrete and continuous inputs, and is shown to be an unbiased Shapley Value estimator with a control variate for variance reduction.

Problem

Feature attribution under black-box (query-only) access is less precise than gradient-based methods, and existing gradient-estimation explainers (GEEX) only work for continuous inputs such as images, not discrete inputs like text.

Approach

Define a search distribution over binary feature masks parameterized by proxy variables in [0,1]^p, estimate proxy gradients via score-function estimator, integrate along the diagonal straight-line path in proxy space, yielding an attribution that is exactly Shapley values without computing marginal contributions; add a control variate based on the number of present features.

Key results

• On text classification, the control-variate GEFA (~GEFA) achieves nAOPC 0.7366 on Amazon with BERT versus 0.6677 for IG and 0.6014 for KernelSHAP.
• On ImageNet with InceptionV3, ~GEFA reaches nAOPC 0.8747, close to the best white-box IG at 0.8805, and above GEEX at 0.7952.
• Theory shows GEFA's attributions are exactly Shapley values and equal Integrated Gradients when both follow the same edge path.

Measured effects reported in the paper

Surprising findings

• The gradient-free-of-baseline black-box GEFA outperforms white-box IG on text classifiers, which the authors attribute to the difficulty of choosing an embedding-space baseline rather than a superiority of black-box methods.
• Vanilla gradients perform at the level of random feature removal on text classifiers.

Limitations

• Evaluation via deletion (nAOPC) is an indirect proxy for explanation quality and may be affected by out-of-distribution deletion artifacts.
• Control variate benefit relies on Assumption 3 (correlation between number of present features and outcome), which may fail for context-dependent text.
• Only next-token attribution for LLMs is explored; generative sequence explanation is left open.

Experimental setup

Datasets: Amazon Review Polarity SST-2 QNLI ImageNet

Benchmarks: Amazon Review Polarity SST-2 QNLI ImageNet

Metrics: normalized AOPC (nAOPC)

Baselines: Vanilla Gradient Integrated Gradients KernelSHAP PartitionSHAP GEEX Random removal

Builds on: Integrated Gradients GEEX KernelSHAP score-function gradient estimator

Closest papers by meaning

• On Gradient-like Explanation under a Black-box Setting: When Black-box Explanations Become as Good as White-box (ICML 2024 · similarity 0.93)
• Provably Better Explanations with Optimized Aggregation of Feature Attributions (ICML 2024 · similarity 0.87)
• Selective Explanations (NeurIPS 2024 · similarity 0.86)
• Towards Fine-Grained Interpretability: Counterfactual Explanations for Misclassification with Saliency Partition (CVPR 2025 · similarity 0.85)
• Prospector Heads: Generalized Feature Attribution for Large Models & Data (ICML 2024 · similarity 0.85)
• Less is More: Fewer Interpretable Region via Submodular Subset Selection (ICLR 2024 · similarity 0.85)