Commute Graph Neural Networks
Authors: Wei Zhuo, Han Yu, Guang Tan, Xiaoxiao Li
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments to evaluate the effectiveness of CGNN on eight digraph datasets. Experimental details and data statistics are provided in Appendix C.1 and Appendix C.2. Table 1 reports the node classification results across eight digraph datasets. Our method CGNN achieves new state-of-the-art results on 6 out of 8 datasets, and comparable results on the remaining ones, validating the superiority of CGNN. Figure 4 compares the accuracy of different models along with running times. Section 5.3 Component Analysis |
| Researcher Affiliation | Academia | Wei Zhuo 1 2 Han Yu 2 Guang Tan 1 Xiaoxiao Li 3 4 Most of this work was done when Wei Zhuo <EMAIL> was with Shenzhen Campus of Sun Yat-sen University. 1Shenzhen Campus of Sun Yat-sen University, China 2Nanyang Technological University, Singapore 3The University of British Columbia, Canada 4Vector Institute, Canada. Correspondence to: Guang Tan <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 CGNN Input: Digraph G = (V, E, X); Depth L; Hidden size d ; Number of classes K Output: Logits ˆY RN K |
| Open Source Code | No | The paper does not explicitly state that the source code for the methodology is available, nor does it provide a link to a code repository. |
| Open Datasets | Yes | The datasets used in Section 5 are Squirrel, Chameleon (Rozemberczki et al., 2021), Citeseer (Sen et al., 2008), Cora ML (Bojchevski & G unnemann, 2017), AM-Photo (Shchur et al., 2018), Snap-Patents, Roman-Empire, and Arxiv-Year (Rossi et al., 2023). |
| Dataset Splits | Yes | For Squirrel and Chameleon, we use 10 public splits (48%/32%/20% for training/validation/testing) provided by (Pei et al., 2019). For the remaining datasets, we adopt the same splits as (Tong et al., 2020a; 2021), which chooses 20 nodes per class for the training set, 500 for the validation set, and allocates the rest to the test set. |
| Hardware Specification | Yes | We conduct our experiments on 2 Intel Xeon Gold 5215 CPUs and 1 NVIDIA Ge Force RTX 3090 GPU. |
| Software Dependencies | No | The paper does not provide specific software dependencies (e.g., library or solver names with version numbers) for its implementation. |
| Experiment Setup | Yes | We utilize the randomized truncated SVD algorithm for computing the Moore-Penrose pseudoinverse of matrix R, setting the required rank q to 5 for all datasets. The learning rate lr is selected from {0.01, 0.005}, and the weight decay wd from {0, 5e 5, 5e 4}. In the model architecture, the number of layers L vary among {1, 2, 3, 4, 5} and the dimension d is selected from {32, 64, 128, 256, 512}. The comprehensive hyperparameter configurations for CGNN are detailed in Table 8. |