Randomized Schur Complement Views for Graph Contrastive Learning

Authors: Vignesh Kothapalli

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on node and graph classification benchmarks demonstrate that our technique consistently outperforms predefined and adaptive augmentation approaches to achieve state-of-the-art results.
Researcher Affiliation Academia 1Courant Institute of Mathematical Sciences, New York University, New York, USA.
Pseudocode Yes The paper contains structured pseudocode blocks: "Algorithm 1 r Lap" and "Algorithm 2 A generalized GCL framework with r Lap."
Open Source Code Yes The code is available at: https://github.com/kvignesh1420/rlap
Open Datasets Yes We experiment on 10 widely used node and graph classification datasets. For node classification, we experiment on CORA (Mc Callum et al., 2000), PUBMED (Sen et al., 2008), AMAZON-PHOTO, COAUTHOR-CS and COAUTHOR-PHY (Shchur et al., 2018). For graph classification, we use the MUTAG (Debnath et al., 1991), PROTEINS-full (PROTEINS) (Borgwardt et al., 2005), IMDB-BINARY, IMDB-MULTI (Yanardag & Vishwanathan, 2015) and NCI1 (Wale et al., 2008) datasets. All the datasets are available in Py G and can be imported directly.
Dataset Splits Yes For testing, we follow a linear evaluation protocol (Velickovic et al., 2019; Zhu et al., 2021b) where the embeddings learned by the encoders in unsupervised settings are classified using a logistic regression classifier with a train, validation and test split of 10%, 10%, 80% respectively.
Hardware Specification Yes We conduct experiments on a virtual machine with 8 Intel(R) Xeon(R) Platinum 8268 CPUs, 24GB of RAM and 1 Quadro RTX 8000 GPU with 32GB of allocated memory.
Software Dependencies Yes For reproducible experiments, we leverage the open-source Py GCL framework by Zhu et al. (2021b) along with Py Torch 1.12.1 (Paszke et al., 2019) and Py Torch-Geometric (Py G) 2.1.0 (Fey & Lenssen, 2019).
Experiment Setup Yes The perturbation ratios γ1, γ2 are selected from {0.0, 0.1, 0.2, 0.3, 0.4, 0.5} to prevent excessive graph corruption. The number of layers in the encoder are chosen from {2, 4, 8}, hidden feature dimensions are chosen from {128, 256, 512}, learning rate is chosen from {10-2, 10-3, 10-4}, weight decay is set to 10-5 and the maximum epoch count is set to 2000. An early stopping strategy with a patience interval of 50 epochs with respect to contrastive loss is also employed. Finally, we choose the Adam optimizer for learning the embeddings. The selected hyper-parameters based on a grid search are reported in Table 12, 13, 14, 15 and can be reproduced with the provided code and instructions.