Generating In-Distribution Proxy Graphs for Explaining Graph Neural Networks
Authors: Zhuomin Chen, Jiaxing Zhang, Jingchao Ni, Xiaoting Li, Yuchen Bian, Md Mezbahul Islam, Ananda Mondal, Hua Wei, Dongsheng Luo
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical evaluations across various datasets demonstrate our method achieves more accurate explanations for GNNs. 6. Experiments |
| Researcher Affiliation | Collaboration | 1Knight Foundation School of Computing and Information Sciences, Florida International University, Miami, USA 2New Jersey Institute of Technology, Newark, USA 3Department of Computer Science, University of Houston, Houston, USA 4Visa Research, USA 5Amazon Search A9, USA 6School of Computing and Augmented Intelligence, Arizona State University, Tempe, USA. |
| Pseudocode | Yes | Algorithm 1 Algorithm of Proxy Explainer (Appendix B) |
| Open Source Code | Yes | The code is available at https://github.com/realMoana/ProxyExplainer. |
| Open Datasets | Yes | MUTAG (Kazius et al., 2005), Benzene (Sanchez-Lengeling et al., 2020),Alkane Carbonyl (Sanchez-Lengeling et al., 2020), and Fluoride Carbonyl (Sanchez-Lengeling et al., 2020), along with two synthetic datasets: BA-2motifs (Luo et al., 2020) and BA3motifs (Chen et al., 2023b). |
| Dataset Splits | No | The paper mentions 'Val Acc' in Table 7, indicating a validation set was used, but does not provide specific details on the dataset split percentages, sample counts, or methodology used to create the validation set, which are needed for full reproducibility. |
| Hardware Specification | Yes | We conduct all experiments on a Linux machine with 8 Nvidia A100-PCIE GPUs, each with 40GB of memory. |
| Software Dependencies | Yes | The CUDA version is 12.4 and the driver version is 550.54.15. We use Python 3.9 and Torch 2.0.1 in our project. |
| Experiment Setup | Yes | We use the Adam optimizer (Kingma & Ba, 2014) with the inclusion of a weight decay 5e 4. We vary λ from 0.01 to 1.0. For D, we vary it among {32, 64, 128, 256, 512, 1024}. M is a hyper-parameter determined by grid search. |