Understanding and Extending Subgraph GNNs by Rethinking Their Symmetries
Authors: Fabrizio Frasca, Beatrice Bevilacqua, Michael Bronstein, Haggai Maron
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, we design a novel Subgraph GNN dubbed SUN, which theoretically unifies previous architectures while providing better empirical performance on multiple benchmarks. . . . Our code is also available. . . We experimentally validate the effectiveness of one Re IGN(2) instantiation, comparing SUN to previously proposed Subgraph GNNs8. We seek to verify whether its theoretical representational power practically enables superior accuracy in expressiveness tasks and real-world benchmarks. Concurrently, we pay attention to the generalisation ability of models in comparison. SUN layers are less constrained in their weight sharing pattern, resulting in a more complex model. As this is traditionally associated with inferior generalisation abilities in low data regimes, we deem it important to additionally assess this aspect. Our code is also available.9 Synthetic. Counting substructures and regressing graph topological features are notoriously hard tasks for GNNs [12, 17, 13]. We test the representational ability of SUN on common benchmarks of this kind [12, 13]. Table 1 reports results on the substructure counting suite, on which SUN attains state-of-the-art results in 3 out of 4 tasks. Additional results on the regression of global, structural properties are reported in Appendix G. Real-world. On the molecular ZINC-12k benchmark (constrained solubility regression) [50, 22, 16], SUN exhibits best performance amongst all domain-agnostic GNNs under the 500k parameter budget, including other Subgraph GNNs (see Table 1). A similar trend is observed on the large-scale Molhiv dataset from the OGB [23] (inhibition of HIV replication). Results are in Table 2. |
| Researcher Affiliation | Collaboration | Fabrizio Frasca Imperial College London & Twitter ffrasca@twitter.com Beatrice Bevilacqua Purdue University bbevilac@purdue.edu Michael M. Bronstein University of Oxford & Twitter mbronstein@twitter.com Haggai Maron NVIDIA Research hmaron@nvidia.com |
| Pseudocode | No | No pseudocode or algorithm block was explicitly labeled or formatted as such in the provided text. |
| Open Source Code | Yes | Our code is also available.9 https://github.com/beabevi/SUN |
| Open Datasets | Yes | On the molecular ZINC-12k benchmark (constrained solubility regression) [50, 22, 16], SUN exhibits best performance amongst all domain-agnostic GNNs under the 500k parameter budget, including other Subgraph GNNs (see Table 1). A similar trend is observed on the large-scale Molhiv dataset from the OGB [23] (inhibition of HIV replication). Results are in Table 2. . . . We experiment on smaller-scale TUDatasets [37] in Appendix G, where we also compare selection policies. |
| Dataset Splits | Yes | Each architecture is selected by tuning the hyperparameters with the entire training and validation sets. We run this experiment on the 4-cycle counting task and the real-world ZINC-12k. |
| Hardware Specification | No | The provided text indicates that hardware specifications are in Appendix G, but Appendix G itself was not provided. Therefore, specific hardware details are not available in the given text. |
| Software Dependencies | No | The paper mentions code availability but does not specify any software dependencies with version numbers in the provided text. |
| Experiment Setup | No | The paper states that training details and hyperparameters are in Appendix G. Since Appendix G was not provided, specific experimental setup details are not available in the given text. |