Learning Divergence Fields for Shift-Robust Graph Representations
Authors: Qitian Wu, Fan Nie, Chenxiao Yang, Junchi Yan
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the practical efficacy of our models on diverse experimental datasets. The results show that the proposed models can effectively handle various distribution shifts. The goal of our experiments is to evaluate the generalization ability of the model under distribution shifts with interdependent data. |
| Researcher Affiliation | Academia | Qitian Wu 1 Fan Nie 1 Chenxiao Yang 1 Junchi Yan 1 School of Artificial Intelligence & Department of Computer Science and Engineering & Mo E Lab of AI, Shanghai Jiao Tong University, Shanghai, China. Correspondence to: Junchi Yan <yanjunchi@sjtu.edu.cn>. |
| Pseudocode | No | No pseudocode or algorithm blocks explicitly labeled as 'Pseudocode' or 'Algorithm' are present. |
| Open Source Code | Yes | Source codes are available at https: //github.com/fannie1208/GLIND. |
| Open Datasets | Yes | We adopt five datasets Twitch, Arxiv, DPPIN, STL and CIFAR for evaluation. Detailed information about these datasets are deferred to Appendix B. Twitch is a multi-graph dataset (Rozemberczki & Sarkar, 2021) ... Arxiv is a temporal network (Hu et al., 2020b) ... DPPIN consists of multiple datasets of biological protein interactions (Fu & He, 2022) ... STL-10 is an image dataset ... CIFAR-10 is another image dataset. |
| Dataset Splits | Yes | In particular, we use the nodes from three subgraphs as training data (where we hold out 25% for validation), and the nodes from the other three subgraphs as testing data. We use the proteins of four datasets for training, one dataset for validation and the other seven datasets for testing. Then we use the training instances with the structures of the first three domains as training data (where we hold out 25% for validation), and the testing instances with the structures of the other three domains as testing data. |
| Hardware Specification | Yes | All of our experiments are run on a Tesla V100 with 16 GB memory. |
| Software Dependencies | Yes | Our implementation is based on Py Torch 1.9.0 and Py Torch Geometric 2.0.3. |
| Experiment Setup | Yes | We adopt Adam with weight decay for training. We set a fixed training budget with 1000 epochs for DPPIN and 500 epochs for other datasets. The searching spaces for all the hyper-parameters are as follows. Number of message passing layers L: [2, 3, 4, 5]. Hidden dimension d: [32, 64, 128]. Dropout ratio: [0.0, 0.1, 0.2, 0.5]. Learning rate: [0.001, 0.005, 0.01, 0.02]. Weight decay: [0, 5e-5, 5e-4, 1e-3]. Number of diffusivity hypothesis K: [3, 4, 5, 10]. Gumbel Softmax temperature τ: [1, 2, 3, 5]. |