GNNs Also Deserve Editing, and They Need It More Than Once
Authors: Shaochen Zhong, Duy Le, Zirui Liu, Zhimeng Jiang, Andrew Ye, Jiamu Zhang, Jiayi Yuan, Kaixiong Zhou, Zhaozhuo Xu, Jing Ma, Shuai Xu, Vipin Chaudhary, Xia Hu
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We propose Sequentially Editable Graph Neural Networks, or SEED-GNN... We find that our method performs vastly better, is more scalable, and remains generalized in contrast to all existing graph-editing techniques under all evaluations we conducted (and we evaluate plenty comprehensively). Additionally, we formally frame the task paradigm of GNN editing and hope to inspire future research in this crucial but currently overlooked field. Please refer to our Git Hub repository for code and checkpoints. |
| Researcher Affiliation | Academia | 1Rice University 2Case Western Reserve University 3Texas A&M University 4North Carolina State University 5Stevens Institute of Technology. |
| Pseudocode | Yes | Algorithm 1 SEED-GNN Training Sample Mix-up Selection ( 5.2) |
| Open Source Code | Yes | Please refer to our Git Hub repository for code and checkpoints. |
| Open Datasets | Yes | To assess the effectiveness of SEED-GNN, we selected seven benchmark datasets from various fields, including common small-scale and large-scale graph datasets like Cora (Mc Callum et al., 2000), Amazon Photo (Shchur et al., 2018), Coauthor CS (Shchur et al., 2018), Amazon Computers (Shchur et al., 2018), ogbn-arxiv, and ogbn-products (Hu et al., 2020). Since model editing is aimed to address high-profile failure cases, we additionally provide results on Yelp CHI (Rayana & Akoglu, 2015), a real-word collected fraud detection dataset. |
| Dataset Splits | Yes | In an inductive setting, we should have three disjoint graphs Dtrain, Dval, and Dtest (Hamilton et al., 2017). In practice, it is recommended to have the edited targets {e1, ..., en} prefetched from the Dval (where M0(ei) = Yei, as all editing targets need to be wrongly predicted by M to be eligible), so that one can easily evaluate the overall test accuracy of a model M(Dtest) before and after edits without further modification. |
| Hardware Specification | No | This work made use of the High Performance Computing Resource in the Core Facility for Advanced Research Computing at Case Western Reserve University. The paper mentions a general 'High Performance Computing Resource' and 'Peak GPU Memory (MB)' in Appendix C, but it does not specify exact GPU or CPU models, processor types, or other detailed hardware configurations used for the experiments. |
| Software Dependencies | No | The paper references various GNN architectures (e.g., GCN, Graph SAGE) and other methods like Adapter-tuning and Lo RA, implying the use of associated software libraries. However, it does not explicitly list any software dependencies with specific version numbers (e.g., 'PyTorch 1.9', 'Python 3.8'). |
| Experiment Setup | Yes | Here, we first report the hyperparameter settings of SEED-GNN in Table 7. We note the following non-SEED-GNN-specific hyperparameter settings are all copied from EGNN by Liu et al. (2023a) for better alignment except LR (where EGNN uses LR=0.01 for most experiments) as we found this smaller LR to be better for our proposed method (Table 11). |