Graph Neural Architecture Search Under Distribution Shifts
Authors: Yijian Qin, Xin Wang, Ziwei Zhang, Pengtao Xie, Wenwu Zhu
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on both synthetic and real-world datasets demonstrate that our proposed GRACES model can adapt to diverse graph structures and achieve state-of-the-art performance for graph classification tasks under distribution shifts. In this section, we report experimental results to verify the effectiveness of our model. We also conduct detailed ablation studies to analyze each of our model components. |
| Researcher Affiliation | Collaboration | 1Department of Computer Science and Technology, Tsinghua University 2THU-Bosch JCML center 3UC San Diego. |
| Pseudocode | Yes | Algorithm 1 The Algorithm Framework of Our Proposed Method |
| Open Source Code | No | Our code will be released at https://github.com/THUMNLab/Auto GL |
| Open Datasets | Yes | Datasets. We adopt both synthetic and real-world datasets for graph-level tasks with distribution shifts. Spurious-Motif (Wu et al., 2022; Ying et al., 2019) is a synthetic dataset... OGBG-Mol* (Hu et al., 2020; Wu et al., 2018) is a set of molecular property prediction datasets. |
| Dataset Splits | Yes | Spurious-Motif 18,000 50/16.7/33.3... OGBG-Mol HIV 41,127 80/10/10... OGBG-Mol SIDER 1,427 80/10/10... OGBG-Mol BACE 1,513 80/10/10 |
| Hardware Specification | Yes | Empirical search time on Spurious-Motif and OGBG-Mol* (NVIDIA Ge Force RTX 3090). |
| Software Dependencies | No | The paper specifies hyper-parameters and training schedules in Appendix B, but does not list specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | We use different learning rate for the three parts of our model. Typically, the learning rate of the self-supervised disentangled encoder is 1.5e-4. The learning rate of the architecture self-customization strategy is 1e-4. Besides, the training procedure of these two parts are consine annealing scheduled. The learning rate of the customized super-network is 2e-3. We initial γ as 0.07 and increase it to 0.5 linearly. In addition, we set β1 = 0.05 and β2 = 0.002. For our method and all baselines, we set the number of layers as 2 in OGBG-Mol HIV and 3 in the other datasets, For all methods, we use edge features in OGBG datasets and virtual node mechanism in OGBG-Mol HIV and OGBG-Mol SIDER. For all methods, we fix the first layer as GIN in Spurious-Motif since some of them cannot deal with constant node features. The hidden dimension in Spurious-Motif is 64 for all baselines and 26 for our method. In OGBG-Mol* datasets, the hidden dimension is 300 for baselines and 128 for our method. |