AutoGCL: Automated Graph Contrastive Learning via Learnable View Generators

Authors: Yihang Yin, Qingzhong Wang, Siyu Huang, Haoyi Xiong, Xiang Zhang8892-8900

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on semi-supervised learning, unsupervised learning, and transfer learning demonstrate the superiority of our Auto GCL framework over the state-of-the-arts in graph contrastive learning.
Researcher Affiliation Collaboration Yihang Yin1, Qingzhong Wang2, Siyu Huang3, Haoyi Xiong2, Xiang Zhang4 1Nanyang Technological University 2Baidu Research 3Harvard University 4The Pennsylvania State University yyin009@e.ntu.edu.sg, wangqingzhong@baidu.com, huang@seas.harvard.edu, xionghaoyi@baidu.com, xzz89@psu.edu
Pseudocode Yes Algorithm 1: Naive training strategy (naive-strategy).", "Algorithm 2: Joint training strategy (joint-strategy).
Open Source Code Yes Our code is available at https://github.com/Somedaywilldo/Auto GCL.
Open Datasets Yes We perform 100 epochs of supervised pre-training on the pre-processed Ch EMBL dataset ((Mayr et al. 2018; Gaulton et al. 2012)), which contains 456K molecules with 1,310 kinds of diverse and extensive biochemical assays." and "We perform semi-supervised graph classification task on TUDataset (Morris et al. 2020)." and "we trained a view generator on MNIST-Superpixel dataset (Monti et al. 2017)".
Dataset Splits Yes We perform a 10-fold cross validation on every dataset. For each fold, we employ 90% of the total data as the unlabeled data for contrastive pre-training, and 10% as the labeled testing data." and "We employ a 10-fold cross validation on each dataset. For each fold, we use 80% of the total data as the unlabeled data, 10% as labeled training data, and 10% as labeled testing data.
Hardware Specification No The paper does not specify the exact hardware used for experiments, such as specific GPU or CPU models.
Software Dependencies No The paper does not provide specific version numbers for software dependencies (e.g., Python, PyTorch, TensorFlow, etc.) that would be required for replication.
Experiment Setup Yes We train the GIN with a batch size of 128 and a learning rate of 0.001. There are 30 epochs of contrastive pre-training under the naive-strategy." and "We use a hidden size of 300 for the classifier, a hidden size of 128 for the view generator. We train the model using a batch size of 256 and a learning rate of 0.001." and "All augmentation ratios are set to 0.2, which is also the default setting in Graph CL.