Simple and Asymmetric Graph Contrastive Learning without Augmentations
Authors: Teng Xiao, Huaisheng Zhu, Zhengyu Chen, Suhang Wang
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show that the simple Graph ACL significantly outperforms state-of-the-art graph contrastive learning and self-supervised learning methods on homophilic and heterophilic graphs. (...) 6 Experiments |
| Researcher Affiliation | Academia | Teng Xiao1 , Huaisheng Zhu1 , Zhengyu Chen2, Suhang Wang1 1The Pennsylvania State University, 2Zhejiang University |
| Pseudocode | No | The paper does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code of Graph ACL is available at https://github.com/tengxiao1/Graph ACL. |
| Open Datasets | Yes | For all datasets, we use the public and standard splits used by cited papers. Detailed descriptions, splits, and one-hop homophily and two-hop monophily statistics of datasets are given in the Appendix C.1. ... For these homophilic datasets, we utilized the process version of them provided by Deep Graph Library [61]. These datasets can be found in https://docs.dgl.ai/en/0.6.x/api/python/ dgl.data.html. ... For Texas, Wisconsin, Cornell, Chameleon, Crocodile, Squirrel and Actor, we use the raw data provided by the Geom-GCN [33] with the standard fixed 10-fold split for our experiment. These datasets can be downloaded from: https://github.com/graphdml-uiuc-jlu/geom-gcn. ... Roman-empire (Roman) [48]... This dataset can be found in https://github.com/yandex-research/heterophilous-graphs. ... Arxiv-year [16]... This dataset can be found in https://github.com/CUAI/Non-Homophily-Large-Scale. |
| Dataset Splits | Yes | For all datasets, we use the public and standard splits used by cited papers. Detailed descriptions, splits, and one-hop homophily and two-hop monophily statistics of datasets are given in the Appendix C.1. ... The public split is used for these datasets, where each class has fixed 20 nodes for training, another fixed 500 nodes and 1000 nodes for validation/test, respectively for evaluation. ... we randomly split the nodes into train/validation/test (10%/10%/80%) sets. ... we also use 10%/10%/80% split for this dataset. ... with the standard fixed 10-fold split for our experiment. ... we utilize the fix 10 random 50%/25%/25% train/validation/test splits. ... 50%/25%/25% train/val/test split is utilized for this dataset. |
| Hardware Specification | Yes | We run experiments on a machine with a NVIDIA RTX A6000 GPU with 49GB of GPU memory. |
| Software Dependencies | No | The paper mentions using the Adam optimizer [62] and Deep Graph Library [61] but does not provide specific version numbers for these or other software dependencies. |
| Experiment Setup | Yes | A small grid search is used to select the best hyperparameter for all methods. In particular, for our Graph ACL, we search λ from {0, 0.90, 0.95, 0.97, 0.99, 0.999, 1}, D from {256, 512, 1024, 2048, 4096, 8192}, τ from {0.25, 0.5, 0.75, 0.99, 1}, and K from {1, 5, 10} when we utilize the negative sampling. We tune the learning rate over {0.001, 0.0005, 0.0001} and weight decay over {0, 0.0001, 0.0003, 0.000001}. |