Rethinking Graph Masked Autoencoders through Alignment and Uniformity
Authors: Liang Wang, Xiang Tao, Qiang Liu, Shu Wu, Liang Wang
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5. Experiments In this section, we conduct experiments to evaluate the effectiveness of AUG-MAE. We analyze it by answering the following questions: RQ1: How does AUG-MAE perform compared with graph self-supervised learning baselines, especially Graph MAE, in various downstream tasks? RQ2: How does each component influence the performance of AUG-MAE? RQ3: How do key hayper-parameters influence the performance of AUG-MAE? RQ4: How does the alignment and uniformity of the representations learned by AUG-MAE, compare with GCL and Graph MAE? |
| Researcher Affiliation | Academia | 1Center for Research on Intelligent Perception and Computing State Key Laboratory of Multimodal Artificial Intelligence Systems Institute of Automation, Chinese Academy of Sciences 2School of Artificial Intelligence, University of Chinese Academy of Sciences |
| Pseudocode | Yes | To help better understand the adversarial training process, we provide the brief pseudo-code of it in Appendix. |
| Open Source Code | Yes | The code is available at: https://github.com/Azure Leon1/AUGMAE. |
| Open Datasets | Yes | We select seven node classification datasets (i.e., Cora, Citeseer (Sen et al. 2008), Pubmed (Namata et al. 2012), Ogbn-arxiv (Hu et al. 2020), PPI, Reddit, and Corafull (Bojchevski and G unnemann 2018)), and six graph classification datasets (i.e., IMDB-B, IMDB-M, PROTEINS, COLLAB, MUTAG, and REDDIT-B (Morris et al. 2020)). |
| Dataset Splits | No | The paper mentions using several benchmark datasets (e.g., Cora, Citeseer, Pubmed, Ogbn-arxiv) and states 'Detailed evaluation setups can be found in Appendix.' However, the main text does not provide specific train/validation/test split percentages, absolute sample counts for each split, or explicit methodology for generating splits, which are needed to reproduce the data partitioning without referring to external appendices or prior work. |
| Hardware Specification | No | The paper does not provide any specific details regarding the hardware (e.g., GPU models, CPU types, memory specifications) used to conduct the experiments. |
| Software Dependencies | No | The paper does not provide specific version numbers for any software dependencies, such as programming languages, libraries, or frameworks (e.g., Python version, PyTorch/TensorFlow version, CUDA version) that would be needed for replication. |
| Experiment Setup | Yes | Fig. 3 shows the effect of varied hyper-parameter values, from which we have the following observations. Effect of weight of ratio regularizer λ1. ... Effect of weight of uniformity regularizer λ2. ... Effect of parameters controlling easy-to-hard α0, αT , η. |