Non-convolutional graph neural networks.
Authors: Yuanqing Wang, Kyunghyun Cho
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | On a wide array of real-world nodeand graph-level tasks, we benchmark the performance of RUM to show its utility in social and physical modeling. Next, to thoroughly examine the performance of RUM, we challenge it with carefully designed illustrative experiments. Specifically, we ask the following questions in this section, with Q1, Q2, and Q3 already theoretically answered in 4: Q1: Is RUM more expressive than convolutional GNNs? Q2: Does RUM alleviate over-smoothing? Q3: Does RUM alleviate over-squashing? Q4: Is RUM slower with convolutional GNNs? Q5: Is RUM robust? Q6: How does RUM scale up to huge graphs? Q7: What components of RUM are contributing most to the performance of RUM? Real-world benchmark performance. For node classification, we benchmark our model on the popular Planetoid citation datasets [47], as well as the coauthor [50] and co-purchase [51] datasets common in social modeling. Additionally, we hypothesize that RUM, without the smoothing operator, will perform competitively on heterophilic datasets [52] we test this hypothesis. For graph classification, we benchmark on the popular TU dataset [53]. We also test the graph regression performance on molecular datasets in Molecule Net [54] and Open Graph Benchmark [55]. In sum, RUM almost always outperforms, is within the standard deviation of, the state-of-the-art architectures, as shown in Tables 2, 3, 4, 5, as well as in Tables 6, 7, 8, 9 moved to the Appendix due to space constraint. |
| Researcher Affiliation | Academia | Yuanqing Wang Center for Data Science and Simons Center for Computational Physical Chemistry New York University New York, N.Y. 10004 wangyq@wangyq.net Kyunghyun Cho Center for Data Science, New York University and Prescient Design, Genetech New York, N.Y. 10004 kc119@nyu.edu |
| Pseudocode | Yes | Algorithm 1 anonymous experiment |
| Open Source Code | Yes | Code at: https://github.com/yuanqing-wang/rum/tree/main |
| Open Datasets | Yes | For node classification, we benchmark our model on the popular Planetoid citation datasets [47], as well as the coauthor [50] and co-purchase [51] datasets common in social modeling. For graph classification, we benchmark on the popular TU dataset [53]. We also test the graph regression performance on molecular datasets in Molecule Net [54] and Open Graph Benchmark [55]. |
| Dataset Splits | Yes | Hyperparameters. All models are optimized using Adam [75] optimizer and Si LU [76] activation functions. 4 random walk samples are drawn everywhere unless specified. Other hyperparameters learning rate (10 5 10 2), hidden dimension (32 64), L2 regularization strength (10 8 10 2), walk length (3 16), temperature for Lconsistency (0 1), coefficient for Lconsistency (0 1), coefficient for Lself, and dropout probability are tuned using the Ray platform [77] with the default Ax [78] search algorithm with 1000 trails or 24 hours tuning budget on a Nvidia A100 GPU. |
| Hardware Specification | Yes | Hyperparameters. All models are optimized using Adam [75] optimizer and Si LU [76] activation functions. 4 random walk samples are drawn everywhere unless specified. Other hyperparameters learning rate (10 5 10 2), hidden dimension (32 64), L2 regularization strength (10 8 10 2), walk length (3 16), temperature for Lconsistency (0 1), coefficient for Lconsistency (0 1), coefficient for Lself, and dropout probability are tuned using the Ray platform [77] with the default Ax [78] search algorithm with 1000 trails or 24 hours tuning budget on a Nvidia A100 GPU. |
| Software Dependencies | No | Core dependencies of our package include Py Torch [64] and Deep Graph Library [38]. While these are mentioned, specific version numbers are not provided in the text. |
| Experiment Setup | Yes | Hyperparameters. All models are optimized using Adam [75] optimizer and Si LU [76] activation functions. 4 random walk samples are drawn everywhere unless specified. Other hyperparameters learning rate (10 5 10 2), hidden dimension (32 64), L2 regularization strength (10 8 10 2), walk length (3 16), temperature for Lconsistency (0 1), coefficient for Lconsistency (0 1), coefficient for Lself, and dropout probability are tuned using the Ray platform [77] with the default Ax [78] search algorithm with 1000 trails or 24 hours tuning budget on a Nvidia A100 GPU. |