Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
SwinGNN: Rethinking Permutation Invariance in Diffusion Models for Graph Generation
Authors: Qi Yan, Zhengyang Liang, Yang Song, Renjie Liao, Lele Wang
TMLR 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on synthetic and real-world protein and molecule datasets show that Swin GNN outperforms existing methods by a substantial margin on most metrics. Our code is released at https://github.com/qiyan98/Swin GNN. 6 EXPERIMENTS We now empirically verify the effectiveness of our model on synthetic and real-world graph datasets including molecule generation with node and edge attributes. |
| Researcher Affiliation | Collaboration | Qi Yan EMAIL University of British Columbia Vector Institute for AI Zhengyang Liang EMAIL Tongji University Yang Song EMAIL Open AI Renjie Liao EMAIL University of British Columbia Vector Institute for AI Canada CIFAR AI Chair Lele Wang EMAIL University of British Columbia |
| Pseudocode | Yes | Algorithm 1 Sampler w. 2nd-order correction. |
| Open Source Code | Yes | Extensive experiments on synthetic and real-world protein and molecule datasets show that Swin GNN outperforms existing methods by a substantial margin on most metrics. Our code is released at https://github.com/qiyan98/Swin GNN. |
| Open Datasets | Yes | Experiment Setup. We consider the following synthetic and real-world graph datasets: (1) Ego-small: 200 small ego graphs from Citeseer dataset (Sen et al., 2008), (2) Community-small: 100 random graphs generated by Erdős Rényi model (Erdös & Rényi, 1959) consisting of two equal-sized communities, (3) Grid: 100 random 2D grid graphs with |V| [100, 400], (4) DD protein dataset (Dobson & Doig, 2003), (5) QM9 dataset (Ramakrishnan et al., 2014), (6) ZINC250k dataset (Irwin et al., 2012). |
| Dataset Splits | Yes | For synthetic and real-world datasets (1-4), we follow the same setup in Liao et al. (2019); You et al. (2018) and apply random split to use 80% of the graphs for training and the rest 20% for testing. In evaluation, we generate the same number of graphs as the test set to compute the maximum mean discrepancy (MMD) of statistics like node degrees, clustering coefficients, and orbit counts. |
| Hardware Specification | Yes | Table 4: Comparison of running time using one NVIDIA RTX 3090 (24 GB) GPU. |
| Software Dependencies | No | The paper mentions "torch.no_grad()" which implies PyTorch but no version. It also mentions "networkx(Hagberg et al., 2008)" but without a specific version number. Therefore, specific ancillary software versions are not provided. |
| Experiment Setup | Yes | Table 6: Architecture details of the proposed Swin GNN and major baselines. The same hyper-parameters are employed for both Swin GNN and Swin GNN-L, barring specific exceptions outlined in the table. The UNet is adopted from Dhariwal & Nichol (2021) with their hyperparameters for the Image Net-64 dataset. Table 7: Training and sampling hyperparameters in the diffusion process. |