On Which Nodes Does GCN Fail? Enhancing GCN From the Node Perspective
Authors: Jincheng Huang, Jialie Shen, Xiaoshuang Shi, Xiaofeng Zhu
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments verify the superiority of the proposed method and demonstrate that current advanced GCNs are improvements specifically on OOC nodes; the remaining nodes under GCN s control (UC nodes) are already optimally represented by vanilla GCN on most datasets. [...] 5. Empirically Studies |
| Researcher Affiliation | Academia | 1School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu, China 2Department of Computer Science, City, University of London, London, United Kingdom 3Sichuan Artificial Intelligence Research Institute, Yibin, China 4Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China, Shenzhen, China. |
| Pseudocode | No | The paper describes algorithms like "Label-Feature Smoothing Alignment Algorithm" but does not present them in a structured pseudocode or algorithm block format. |
| Open Source Code | No | The paper does not contain any explicit statement about releasing source code or a link to a code repository. |
| Open Datasets | Yes | We evaluate the effectiveness of the proposed method on seven datasets based on citation, co-authorship, and co-purchase graphs for semi-supervised node classification tasks; including Cora, Citeseer, Pubmed (Sen et al., 2008), Coauthor CS, Coauthor Physics, Amazon Computers and Amazon Photo (Shchur et al., 2018). |
| Dataset Splits | Yes | We follow the evaluation protocol and split of (Kipf & Welling, 2017) on the Citation Network dataset (i.e., 20 per class for train, 500 nodes for valid, 1000 nodes for test) and follow (Shchur et al., 2018; Liu et al., 2020) on the co-authorship and co-purchase datasets(i.e., 20 per class for train, 30 per class for valid, remain nodes for test). |
| Hardware Specification | Yes | We conduct all experiments on a server with Nvidia RTX 4090 (24GB memory each) and conduct each experiment on ten random seeds and report the average results. |
| Software Dependencies | No | The paper mentions using "Adam optimization" but does not specify version numbers for any programming languages, libraries, or frameworks used in the implementation (e.g., Python, PyTorch, TensorFlow, scikit-learn versions are not provided). |
| Experiment Setup | Yes | In the proposed method, we optimize all parameters by Adam optimization (Kingma & Ba, 2015) with the learning rate 0.01 except for Pubmed is 0.2, and set the weight decay as 0.0005 for all datasets. Moreover, we set the number of model layers in the range of {2,3,4}, set the dropout in the range of {0.5, 0.6}, and set the size of the hidden unit in the range of {4, 8, 16, 62, 64}. We set τ in from 0 to 1 at intervals of 0.1 and β in from 0 to 1.5 at intervals of 0.1. The best hyperparameters for each dataset can be found in Appendix B.5. |