GEFA: A General Feature Attribution Framework Using Proxy Gradient Estimation
Authors: Yi Cai, Thibaud Ardoin, Gerhard Wunder
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Compared to traditional sampling-based Shapley Value estimators, GEFA avoids potential information waste sourced from computing marginal contributions, thereby improving explanation quality, as demonstrated in quantitative evaluations across various settings.5. Experiments |
| Researcher Affiliation | Academia | 1Department of Mathematics and Computer Science, Freie Universit at Berlin, Berlin, Germany. Correspondence to: Yi Cai <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 GEFA Explanation Scheme Algorithm 2 Smoothing Enhanced Mask Sampling |
| Open Source Code | Yes | 2Code is available at: https://github.com/caiy0220/GEFA |
| Open Datasets | Yes | Three datasets are adopted for text classification tasks: Amazon Review Polarity (Mc Auley & Leskovec, 2013), STS-2, and QNLI (Wang et al., 2019). The image classification task is set up with Image Net (Russakovsky et al., 2015) |
| Dataset Splits | Yes | Without losing generality, we adopted a lightweight model for image classification and downsampled the dataset into 2000/400/400 partitions for training, validation, and test sets to ensure feasibility and efficiency. |
| Hardware Specification | Yes | Processor: Intel i9-10980XE, 18 cores Memory: 32GB DDR4 GPU: NVIDIA RTX A5500, 24GB |
| Software Dependencies | Yes | The primary packages were Numpy 1.26.4, Py Torch of version 2.5.0, and Torchvision 0.20.0. The CUDA version was 12.2 for GPU support. |
| Experiment Setup | Yes | For all test cases, the query budget for the black-box explainers is 500, given the relatively smaller feature space; the interpolation step for IG is set to 50. The query budget of the black-box approaches is increased to 5000 due to the considerably larger input feature spaces, which are 299 299 and 224 224 for Inception V3 and Vi T, respectively. |