Graph Contrastive Invariant Learning from the Causal Perspective

Authors: Yanhu Mo, Xiao Wang, Shaohua Fan, Chuan Shi

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate the effectiveness of our approach on node classification tasks.
Researcher Affiliation Academia Yanhu Mo1, Xiao Wang2 , Shaohua Fan3,4, Chuan Shi1* 1Beijing University of Posts and Telecommunications 2 Beihang University 3 Tsinghua University 4 Key Laboratory of Big Data Artificial Intelligence in Transportation, Ministry of Education(Beijing Jiaotong University)
Pseudocode No The paper describes the pipeline and components of GCL and GCIL in narrative text, but it does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks or figures.
Open Source Code Yes GCIL1 1https://github.com/BUPT-GAMMA/GCIL
Open Datasets Yes To evaluate our method, we consider five commonly used node classification benchmark datasets from the previous works (Velickovic et al. 2019; Mernyei and Cangea 2020; Liu et al. 2022a), including Cora, Citeseer, Pubmed, Wiki CS, and Flickr. The statistic of these datasets is summarized in Table 1.
Dataset Splits Yes We adopt the public splits for Cora, Citeseer, Pubmed, and Flickr, where the training set contains 20 nodes per class, 500 nodes for validation, and 1000 for testing. For the Wiki-CS dataset, we evaluate the models on the public splits provided in (Mernyei and Cangea 2020).
Hardware Specification No The paper vaguely mentions 'on a 24GB GPU' in a footnote for an OOM error, but provides no specific details about the hardware (e.g., GPU model, CPU, memory, number of GPUs) used for running the experiments.
Software Dependencies No The paper does not explicitly list specific software dependencies with their version numbers (e.g., 'Python 3.8, PyTorch 1.9, CUDA 11.1').
Experiment Setup Yes Parameter Settings. We set the dimensions of all datasets to 512, with a learning rate of 0.0002 for the Flickr dataset and 0.001 for all other datasets. The weight decay for all datasets is 0.0001. Additional details of parameter settings are presented in Appendix C. and where α, β, and γ are hyper-parameters controlling the importance of each term in the loss. The λ represents the desired standard deviation of the dimensions.