Graph Structure Extrapolation for Out-of-Distribution Generalization

Authors: Xiner Li, Shurui Gui, Youzhi Luo, Shuiwang Ji

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5. Experimental Studies
Researcher Affiliation Academia 1Department of Computer Science and Engineering, Texas A&M University, Texas, USA.
Pseudocode No The paper describes procedures in text and uses mathematical formulations but does not provide clearly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not contain an explicit statement about releasing source code or a link to a code repository for the methodology described.
Open Datasets Yes We adopt 5 datasets from the GOOD benchmark (Gui et al., 2022a), HIV-size, HIV-scaffold, SST2-length, Motif-size, and Motif-base, where -" denotes the shift domain. We construct another natural language dataset Twitter-length (Yuan et al., 2020) following the OOD split of GOOD. Additionally, we adopt protein dataset DD-size and molecular dataset NCI1-size following Bevilacqua et al. (2021).
Dataset Splits Yes For all experiments, we select the best checkpoints for OOD tests according to results on OOD validation sets; ID validation and ID test are also used for comparison if available.
Hardware Specification Yes For computation, we generally use one NVIDIA Ge Force RTX 2080 Ti for each single experiment.
Software Dependencies No The paper mentions software components like "GIN-Virtual Node", "GIN", "Graph SAINT", "GCN", and "Adam optimizer" but does not provide specific version numbers for any of them.
Experiment Setup Yes For all the experiments, we use the Adam optimizer, with a weight decay tuned from the set {0, 1e-2, 1e-3, 1e-4} and a dropout rate of 0.5. The number of convolutional layers in GNN models for each dataset is tuned from the set {3, 5}. We use mean global pooling and the RELU activation function, and the dimension of the hidden layer is 300. We select the maximum number of epochs from {100, 200, 500}, the initial learning rate from {1e-3, 3e-3, 5e-3, 1e-4}, and the batch size from {32, 64, 128} for graph-level and {1024, 4096} for node-level tasks. All models are trained to converge in the training process.