Graph Parsing Networks
Authors: Yunchong Song, Siyuan Huang, Xinbing Wang, Chenghu Zhou, Zhouhan Lin
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on standard benchmarks demonstrate that GPN outperforms state-of-the-art graph pooling methods in graph classification tasks while being able to achieve competitive performance in node classification tasks. We also conduct a graph reconstruction task to show GPN s ability to preserve node information and measure both memory and time efficiency through relevant tests. |
| Researcher Affiliation | Academia | Yunchong Song1, Siyuan Huang1, Xinbing Wang1, Chenghu Zhou2, Zhouhan Lin1 1Shanghai Jiao Tong University, 2Chinese Academy of Sciences |
| Pseudocode | Yes | Algorithm 1 Graph parsing algorithm A |
| Open Source Code | Yes | Code is available at https://github.com/LUMIA-Group/Graph Parsing Networks |
| Open Datasets | Yes | We evaluate our model on five widely-used graph classification benchmarks from TUDatasets (Morris et al., 2020): DD, PROTEINS, NCI1, NCI109, and FRANKENSTEIN. Additionally, we evaluate our model s performance at scale on ogbg-molpcba, one of the largest graph classification datasets in OGB (Hu et al., 2020)... We conduct node classification experiments under full supervision for four datasets: Cora, Cite Seer, and Pub Med from Sen et al. (2008), and Film from Pei et al. (2020). |
| Dataset Splits | Yes | We adopt 10-fold cross-validation with 20 random seeds for both model initialization and fold generation. This means there are 200 runs in total behind each data point, ensuring a fair comparison. For all datasets, we use an early stopping strategy to avoid overfitting. We use validation loss as the criterion the early stopping. |
| Hardware Specification | Yes | We use NVIDIA Ge Force RTX 3090 and NVIDIA Ge Force RTX 4090 as the hardware environment |
| Software Dependencies | No | The paper states: 'We use Py Torch and Py Torch Geometric (Fey & Lenssen, 2019) as our software environment.' However, it does not specify version numbers for these software components, which is required for a reproducible description of ancillary software. |
| Experiment Setup | Yes | The Table 5 gives the best hyperparameters used in the 5 graph classification datasets. For the sake of brevity, we use abbreviations in the table. LGNN represents number of GCN layers inside the GNN block, LDeep Sets represents number of MLP layers of Deep Sets, LMLP represents number of layers of the MLP for edge score computation, Dropout represents dropout in the MLP, Drop Edge represents the ratio of the edge dropout when perform parsing, lr represents learning rate, H represents hidden channel, B represents batch size. We use an early stopping strategy to avoid overfitting. We use validation loss as the criterion the early stopping. Also, the maximum number of epochs is set to 500 for all datasets. |