Towards Bridging Generalization and Expressivity of Graph Neural Networks
Authors: Shouheng Li, Floris Geerts, Dongwoo Kim, Qing Wang
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through case studies and experiments on real-world datasets, we demonstrate that our theoretical findings align with empirical results, offering a deeper understanding of how expressivity can enhance GNN generalization. 7 EXPERIMENTS Tasks and Datasets We conduct graph classification experiments on six widely used benchmark datasets: ENZYMES, PROTEINS, and MUTAG from the TU dataset collection (Morris et al., 2020a), as well as SIDER and BACE from the molecular dataset collection (Wu et al., 2017). |
| Researcher Affiliation | Academia | 1 School of Computing, Australian National University, Australia 2 Department of Computer Science, University of Antwerp, Belgium 3 CSE & GSAI, POSTECH, South Korea 4 Data61, CSIRO, Australia |
| Pseudocode | Yes | Algorithm 1: An algorithm to compute the bound in Theorem D.2 |
| Open Source Code | Yes | The code implementation is available at https://github.com/seanli3/hom_gen. |
| Open Datasets | Yes | We conduct graph classification experiments on six widely used benchmark datasets: ENZYMES, PROTEINS, and MUTAG from the TU dataset collection (Morris et al., 2020a), as well as SIDER and BACE from the molecular dataset collection (Wu et al., 2017). |
| Dataset Splits | Yes | Each dataset is randomly divided into training and test sets following a 90%/10% split. |
| Hardware Specification | No | The paper does not explicitly describe the hardware used to run its experiments, only mentioning general experimental setup without specific GPU/CPU models or processor types. |
| Software Dependencies | No | The paper does not provide specific ancillary software details with version numbers. It mentions the use of a softmax function, but not specific libraries or frameworks with their versions. |
| Experiment Setup | Yes | Each classification task is trained for 400 epochs, with five independent runs to report the mean and standard deviation of the results. Consistent with the setup in Tang and Liu (2023); Morris et al. (2023); Cong et al. (2021), we eliminate the use of regularization techniques such as dropout and weight decay. A batch size of 128 is utilized, with a learning rate set to 10 3, and the hidden layer dimension fixed at 64. The margin loss function is employed with a margin parameter γ = 1. [...] For all experiments, we set the confidence level δ to 0.1, yielding bounds with high probability. |