Accurate Learning of Graph Representations with Graph Multiset Pooling
Authors: Jinheon Baek, Minki Kang, Sung Ju Hwang
ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experimental results show that GMT significantly outperforms state-of-the-art graph pooling methods on graph classification benchmarks with high memory and time efficiency, and obtains even larger performance gain on graph reconstruction and generation tasks. |
| Researcher Affiliation | Collaboration | Jinheon Baek1 , Minki Kang1 , Sung Ju Hwang1,2 KAIST1, AITRICS2, South Korea |
| Pseudocode | No | The paper describes the architecture and components in text and mathematical equations, but it does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks, nor does it present structured steps in a pseudocode-like format. |
| Open Source Code | Yes | Code is available at https://github.com/Jinheon Baek/GMT |
| Open Datasets | Yes | To validate the proposed Graph Multiset Transformer (GMT) for graph representation learning, we evaluate it on classification, reconstruction and generation tasks of synthetic and real-world graphs. ... Among TU datasets (Morris et al., 2020), we select 6 datasets... Also, we use 4 molecule datasets (HIV, Tox21, Tox Cast, BBBP) from the OGB datasets (Hu et al., 2020)... We further experiment with real-world Molecule Graph, namely ZINC datasets (Irwin et al., 2012)... We use the QM9 dataset (Ramakrishnan et al., 2014) following the original Mol GAN paper (Cao & Kipf, 2018). ...We use the USPTO-50k dataset following the original paper (Dai et al., 2019). |
| Dataset Splits | Yes | For all experiments on TU datasets, we evaluate the model performance with a 10-fold cross validation setting, where the dataset split is based on the conventionally used training/test splits (Niepert et al., 2016; Zhang et al., 2018; Xu et al., 2019), with LIBSVM (Chang & Lin, 2011). In addition, we use the 10 percent of the training data as a validation data following the fair comparison setup (Errica et al., 2020). For all experiments on OGB datasets, we evaluate the model performance following the original training/validation/test dataset splits (Hu et al., 2020). |
| Hardware Specification | No | The paper mentions 'GPU Memory Efficiency' and measures 'forward time, including CPU and GPU', but does not specify any particular hardware models such as GPU or CPU types, or detailed computing cluster specifications. |
| Software Dependencies | No | The paper mentions using specific optimizers like 'Adam optimizer (Kingma & Ba, 2014)' and libraries like 'LIBSVM (Chang & Lin, 2011)', along with 'GCN layers (Kipf & Welling, 2017)', but it does not specify software dependencies with version numbers such as Python, PyTorch, or specific deep learning frameworks. |
| Experiment Setup | Yes | For all experiments on TU datasets except the D&D, the learning rate is set to 5 10 4, hidden size is set to 128, batch size is set to 128, weight decay is set to 1 10 4, and dropout rate is set to 0.5. Since the D&D dataset has a large number of nodes (See Table 1 in the main paper), node clustering methods can not perform clustering operations on large graphs with large batch sizes, such that the hidden size is set to 32, and batch size is set to 10 on the D&D dataset. |