Robust Graph Neural Networks via Unbiased Aggregation
Authors: Zhichao Hou, Ruiqi Feng, Tyler Derr, Xiaorui Liu
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our comprehensive experiments confirm the strong robustness of our proposed model under various scenarios, and the ablation study provides a deep understanding of its advantages. Our code is available at https://github.com/chris-hzc/RUNG. |
| Researcher Affiliation | Academia | Zhichao Hou1 Ruiqi Feng1 Tyler Derr2 Xiaorui Liu1 1North Carolina State University, 2Vanderbilt University |
| Pseudocode | No | The paper describes algorithms using mathematical equations and prose but does not provide structured pseudocode or an algorithm block. |
| Open Source Code | Yes | Our code is available at https://github.com/chris-hzc/RUNG. |
| Open Datasets | Yes | We test our RUNG with the node classification task on two widely used real-world citation networks, Cora ML and Citeseer [29], as well as a large-scale networks Ogbn-Arxiv [30]. |
| Dataset Splits | Yes | We adopt the data split of 10% training, 10% validation, and 80% testing, and report the classification accuracy of the attacked nodes following [5]. |
| Hardware Specification | No | The paper does not provide specific details on the hardware used, such as GPU or CPU models. |
| Software Dependencies | No | The paper mentions libraries/frameworks implicitly through citations but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | The model hyperparameters including learning rate, weight decay, and dropout rate are tuned as in [5]. Other hyperparameters follow the settings in the original papers. RUNG uses an MLP connected to 10 graph aggregation layers following the decoupled GNN architecture of APPNP. ˆλ = 1 / (1 + λ) is tuned in {0.7, 0.8, 0.9}, and γ tuned in {0.5, 1, 2, 3, 5}. We chose the hyperparameter setting that yields the best robustness without a notable impact (smaller than 1%) on the clean accuracy following [35]. |