G-Mixup: Graph Data Augmentation for Graph Classification
Authors: Xiaotian Han, Zhimeng Jiang, Ninghao Liu, Xia Hu
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that G-Mixup substantially improves the generalization and robustness of GNNs. |
| Researcher Affiliation | Academia | 1Department of Computer Science&Engineering, Texas A&M University 2Department of Computer Science, University of Georgia 3Department of Computer Science, Rice University. |
| Pseudocode | Yes | Algorithm 1 Graphon Estimation ... Algorithm 2 G-Mixup ... Algorithm 3 G-Mixup (batch) |
| Open Source Code | Yes | Our code is available at https://github.com/ahxt/g-mixup. |
| Open Datasets | Yes | We evaluate the performance of G-Mixup ... with various datasets and GNN backbones in Section 5.3... Dataset IMDB-B IMDB-M REDD-B REDD-M5 REDD-M12... We experiment on molecular property prediction task (Hu et al., 2020), including ogbg-molhiv, ogbg-molbace, ogbgmolbbbp. |
| Dataset Splits | Yes | We split the dataset into train/val/test data by 7 : 1 : 2. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory specifications) used for running the experiments. |
| Software Dependencies | No | The paper mentions specific algorithms and models like 'Adam optimizer(Kingma & Ba, 2015)', 'GCN (Kipf & Welling, 2017)', 'GIN (Xu et al., 2018)', and refers to implementations like 'https://github.com/pyg-team/pytorch_geometric/blob/1.7.2/examples/gcn2_ppi.py'. However, it does not provide a list of ancillary software dependencies with specific version numbers (e.g., Python version, PyTorch version). |
| Experiment Setup | Yes | For model training, we use the Adam optimizer(Kingma & Ba, 2015). The initial learning rate is 0.01 and will drop the learning rate by half every 100 epochs. The batch size is set to 128. ... For hyperparemeter in G-Mixup, we generate 20% more graphs for training graph. The graphons are estimated based on the training graphs. We use different λ [0.1, 0.2] to mix up the graphon and generate synthetic with different strength of mixing up. |