Diffusion on Graph: Augmentation of Graph Structure for Node Classification
Authors: Yancheng Wang, Changyu Liu, Yingzhen Yang
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on various graph datasets for semi-supervised node classification and graph contrastive learning have been conducted to demonstrate the effectiveness of Do G with low-rank regularization. The code of Do G is available at https://github.com/Statistical-Deep-Learning/Do G. |
| Researcher Affiliation | Academia | Yancheng Wang EMAIL Changyu Liu EMAIL Yingzhen Yang EMAIL School of Computing and Augmented Intelligence Arizona State University |
| Pseudocode | Yes | Algorithm 1 and Algorithm 2 in Section D of the appendix describe the training algorithm of the Do G and the generation process of the augmented graph Gaug in details. |
| Open Source Code | Yes | The code of Do G is available at https://github.com/Statistical-Deep-Learning/Do G. |
| Open Datasets | Yes | We conduct experiments on five public benchmarks that are widely used for node classification on attributed graphs, namely Cora, Citeseer, Pubmed (Sen et al., 2008a), Coauthor CS, and ogbn-arxiv (Hu et al., 2020). Details on the statistics of the dataset are deferred to Table 6 in Section E.1 of the appendix. |
| Dataset Splits | Yes | For all our experiments, we follow the default separation (Shchur et al., 2018; Mernyei & Cangea, 2020; Hu et al., 2020) of training, validation, and test sets on each benchmark. ... We select the values of γ, τ, and β by performing 5-fold cross-validation on 20% of the training data in each dataset. |
| Hardware Specification | Yes | We perform all the experiments in our paper on one NVIDIA Tesla A100 GPU. |
| Software Dependencies | No | The paper mentions using optimizers (Adam, Adam W), MLPs, GAT, but does not provide specific version numbers for any libraries, programming languages (like Python), or frameworks (like PyTorch/TensorFlow/CUDA). |
| Experiment Setup | Yes | We use Adam optimizer with a learning rate of 0.001 for the training of the GAE. The weight decay is set to 1 10 5. ... We use the Adam W optimizer to optimize the LDM with a learning rate of 0.0002 and a weight decay factor of 0.0001. ... The guidance strength of CFG is set to 0.5 in our experiments. A three-layer Multilayer Perceptron (MLP) is used as the denoising model in the LDM, whose hidden dimension is set to 512. We train LDM for 3000 epochs and keep track of the exponential moving average (EMA) of the model during the training with a decay factor of 0.995. ... The value of γ is selected from {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9}. The value of τ is selected from {0.05, 0.1, 0.15, 0.2, 0.25, 0.3, 0.35, 0.4, 0.45, 0.5}. The value of β is selected from {1, 2, 3, 4, 5, 6, 7, 8, 9, 10}. |