Spectral Feature Augmentation for Graph Contrastive Learning and Beyond

Authors: Yifei Zhang, Hao Zhu, Zixing Song, Piotr Koniusz, Irwin King

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on graph/image datasets show that our spectral feature augmentation outperforms baselines, and is complementary with other augmentation strategies and compatible with various contrastive losses.
Researcher Affiliation Collaboration 1The Chinese University of Hong Kong 2Australian National University 3Data61/CSIRO
Pseudocode Yes Algorithm 1: Spectral Feature Augmentation (AF )
Open Source Code No The paper states 'Implementation and evaluation are based on Py GCL (Zhu et al. 2021a): https://github.com/Py GCL/Py GCL.' and 'Implementation/evaluation are based on Solo-learn (da Costa et al. 2022): https://github.com/vturrisi/solo-learn.', which refer to external libraries used. It does not explicitly state that the authors' specific implementation code for the proposed method is openly available, nor does it provide a direct link to their own repository.
Open Datasets Yes Datasets. We use five popular datasets (Zhu et al. 2020, 2021b; Velickovic et al. 2019), including citation networks (Cora, Cite Seer) and social networks (Wiki-CS, Amazon Computers, Amazon-Photo) (Kipf and Welling 2017; Sinha et al. 2015; Mc Auley et al. 2015; Mernyei and Cangea 2020). For graph classification, we use NCI1, PROTEIN and DD (Dobson and Doig 2003; Riesen and Bunke 2008). For image classification we use CIFAR10/100 and Image Net-100 (Deng et al. 2009).
Dataset Splits Yes Graph Datasets are randomly divided into 10%, 10%, 80% for training, validation, and testing.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. It only mentions general terms like 'GCN layers'.
Software Dependencies No The paper mentions using 'Py GCL' and 'Solo-learn' for implementation and evaluation, but does not provide specific version numbers for these libraries or any other software dependencies (e.g., Python, PyTorch, CUDA).
Experiment Setup Yes We use Xavier initialization for the GNN parameters and train the model with Adam optimizer. For node/graph classification, we use 2 GCN layers. The logistic regression classifier is trained with 5, 000 (guaranteed converge). We also use early stopping with a patience of 20 to avoid overfitting. We set the size of the hidden dimension of nodes to from 128 to 512. For the chosen hyper-parameters see Section D.2.