SBGD: Improving Graph Diffusion Generative Model via Stochastic Block Diffusion
Authors: Junwei Su, Shan Wu
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical results show that SBGD achieves significant memory improvements (up to 6 ) while maintaining comparable or even superior graph generation performance relative to state-of-the-art methods. Furthermore, experiments demonstrate that SBGD better generalizes to unseen graph sizes. The significance of SBGD extends beyond being a scalable and effective GDGM; it also exemplifies the principle of modularization in generative modeling, offering a new avenue for exploring generative models by decomposing complex tasks into more manageable components. |
| Researcher Affiliation | Academia | 1School of Computing and Data Science, University of Hong Kong 2School of Resources and Environmental Engineering, Hefei University of Technology. Correspondence to: Junwei Su <EMAIL>, Shan Wu <EMAIL >. |
| Pseudocode | Yes | Pseudo-code for training and sampling is provided in Appendix B, along with additional technical details of the implementation. ... Algorithm 1 SBGD Training Algorithm ... Algorithm 2 SBGD sampling algorithm. |
| Open Source Code | No | The paper does not contain any explicit statements about releasing source code for the methodology, nor does it provide a link to a code repository. It only mentions that pseudocode is provided in the appendix. |
| Open Datasets | Yes | Datasets. We consider five real and synthetic datasets with varying sizes and connectivity levels: Planar-graphs, Contextual Stochastic Block Model(c SBM) (Deshpande et al., 2018), Proteins (Dobson & Doig, 2003), QM9 (Wu et al., 2018), OGBN-Arxiv, and OGBN-Products (Hu et al., 2021). |
| Dataset Splits | No | The paper mentions using 'test and generated graphs' for evaluation but does not specify the splitting methodology (e.g., percentages, counts, or predefined splits) for the datasets used in the experiments. |
| Hardware Specification | Yes | Testbed. Our experiments were conducted on a Dell Power Edge C4140, The key specifications of this server, pertinent to our research, include: CPU: Intel Xeon Gold 6230 processors equipped with 20 cores and 40 threads, GPU: NVIDIA Tesla V100 SXM2 units equipped with 32GB of memory, Memory: An aggregate of 256GB RAM, distributed across eight 32GB RDIMM modules, and Operating System: Ubuntu 18.04LTS |
| Software Dependencies | No | The paper mentions using a 'Graph Transformer' and the 'Adam optimizer', and refers to the 'METIS algorithm' from the 'DGL library', but no specific version numbers are provided for these software components. |
| Experiment Setup | Yes | For training our network, we adopt the widely-used Adam optimizer, tuning only the learning rate as the primary hyperparameter. To determine the optimal values for other hyperparameters in our model, we perform a simple grid search over the following ranges: Number of layers: [2, 4], Hidden dimension: [8, 16, 32, 64, 128, 256], Learning rate: [0.1, 0.05, 0.01, 0.005, 0.001], Diffusion Length T: [50,100,200], Sampling Steps: [100,200,500,1000]. For the variance schedule, we follow the one in (Jo et al., 2022). |