Contrastive Learning Meets Homophily: Two Birds with One Stone
Authors: Dongxiao He, Jitao Zhao, Rui Guo, Zhiyong Feng, Di Jin, Yuxiao Huang, Zhen Wang, Weixiong Zhang
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our extensive empirical results demonstrate that the new method can significantly outperform the existing GCL methods because the former can solve the homophily problem in a self-supervised way with the new group discrimination method used. Finally, we experimentally demonstrate the performance of our Ne Co method on node classification tasks using five heterophily datasets and four commonly used homophily datasets. The Ne Co method significantly improves performance compared to the state-of-the-art baselines. |
| Researcher Affiliation | Academia | 1College of Intelligence and Computing, Tianjin University, Tianjin, China 2Department of Data Science, George Washington University, NW Washington DC, America 3Department of Artificial Intelligence, OPtics and Electro Nics (i OPEN), Northwestern Polytechnical University, Xi an, China 4Department of Health Technology and Informatics, Department of Computing, The Hong Kong Polytechnic University, Hong Kong, China. |
| Pseudocode | No | The paper does not include a figure, block, or section explicitly labeled 'Pseudocode', 'Algorithm', or 'Algorithm X', nor does it present structured steps formatted like code. |
| Open Source Code | No | The paper does not contain an explicit statement about releasing the code for the methodology described, nor does it provide a direct link to a source-code repository. |
| Open Datasets | Yes | The homophily datasets we used are citation networks, including Cora, Citeseer, Pubmed and DBLP, and the heterophily datasets consist of web page networks (Cornell, Texas and Wisconsin), Actor and Wikipedia network Chameleon. All datasets are available in Py Torch Geometric library. |
| Dataset Splits | Yes | For the homophily networks, we adopted the commonly-used 10%/10%/80% nodes for training, validation and testing. We changed the data split for heterophily datasets to 60%/20%/20%, which follows the existing GNN works on heterophily problems since strong homophily datasets contain rich label information so that during message passing. In contrast, the heterophily dataset needs more labels to fine-tune the downstream classifier. |
| Hardware Specification | Yes | All our experiments were performed on the Google Colab platform with a Tesla NVIDIA Tesla P100 (16GB) GPU. |
| Software Dependencies | No | The paper mentions that datasets are available in the 'Py Torch Geometric library' and discusses GCN and Graph SAGE as base encoders, but it does not provide specific version numbers for these or any other software components. |
| Experiment Setup | Yes | Here we study how the different temperature hyperparameter τ and edge drop ratio p affect the training of our proposed Ne Co. We vary the hyperparameter τ within the scope of {0.1,0.3,0.5,1,2,5} to observe the accuracy variation of Ne Co and GRACE on the test set. For the drop ratio p, we modify the regularization term in Equation 5 to λ(|| P i V N(i; θ)||/|E| p) to control the proportion of edges that Ne Co drops. |