Unified Graph Augmentations for Generalized Contrastive Learning on Graphs
Authors: Jiaming Zhuo, Yintong Lu, Hui Ning, Kun Fu, bingxin niu, Dongxiao He, Chuan Wang, Yuanfang Guo, Zhen Wang, Xiaochun Cao, Liang Yang
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Evaluations across various datasets and tasks demonstrate the generality and efficiency of the proposed GOUDA over existing state-of-the-art GCLs. |
| Researcher Affiliation | Academia | 1Hebei Province Key Laboratory of Big Data Calculation, School of Artificial Intelligence, Hebei University of Technology, Tianjin, China 2College of Intelligence and Computing, Tianjin University, Tianjin, China 3School of Computer Science and Technology, Beijing Jiao Tong University, Beijing, China 4School of Computer Science and Engineering, Beihang University, Beijing, China 5School of Artificial Intelligence, OPtics and Electro Nics (i OPEN), School of Cybersecurity, Northwestern Polytechnical University, Xi an, China 6School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University, Shenzhen, China |
| Pseudocode | Yes | Algorithm 1: GOUDA-IF |
| Open Source Code | Yes | We have included complete and executable code within the supplemental material, ensuring the reproducibility of our results. |
| Open Datasets | Yes | We source these datasets from the public repository Py Torch Geometric (Py G) [9]. The datasets can be accessed through the URLs listed below: Cora, Cite Seer, Pub Med: https://github.com/kimiyoung/planetoid/raw/master/data. Wiki-CS: https://github.com/pmernyei/wiki-cs-dataset/raw/master/dataset. Photo, Computers: https://github.com/shchur/gnn-benchmark/raw/master/data/npz/. Physics: https://github.com/shchur/gnn-benchmark/raw/master/data/npz/. IMDB-B, IMDB-M, COLLAB: https://ls11-www.cs.tu-dortmund.de/staff/morris/graph kerneldatasets. |
| Dataset Splits | Yes | For the seven benchmark datasets utilized for node classification tasks (Cora, Citeseer, Pub Med, Wiki CS, Computers, Photo, and Physics), the dataset is divided into training, validation, and testing sets in the ratio of 1:1:8. |
| Hardware Specification | Yes | Table 6: Experimental environment servers. Server 1 OS Linux 5.15.0-82-generic CPU Intel(R) Core(TM) i7-12700K CPU @ 3.6GHz GPU Ge Force RTX 4090. Server 2 OS Linux 5.15.0-100-generic CPU Intel(R) Core(TM) i9-10980XE CPU @ 3.00GHz GPU Ge Force RTX 3090 |
| Software Dependencies | No | The paper mentions software like 'Py Torch Geometric (Py G)' and 'Py G' for implementations but does not specify their version numbers, nor does it list specific versions for other key software dependencies (e.g., Python, PyTorch, CUDA). |
| Experiment Setup | Yes | GOUDA is implemented as two models: GOUDA-IF, which utilizes Info NCE loss, and GOUDABT, which employs Barlow Twins loss. For the node-level tasks, both models are trained using an Adam optimizer with a learning rate of 1e 3 and the weight decay rate from {0, 5e 5, 5e 4}. The dimensions d of node embeddings are selected from {256, 512, 1024, 2048}, and their impact is analyzed in Section 4.2. The hyperparameters β1 and β2 of independence loss are chosen from {1e 4, 1e 3}, while the related hyperparameter γ is selected among {1e 2, 1e 1, 1, 10}. Additionally, for GOUDA-IF, the temperature coefficient τ is selected from {0.2, 0.4, 0.6, 0.8}, while for GOUDA-BT, the hyperparameter λ is set to 1 d. For the graph-level task, the configuration follows Graph CL [48], where the hidden dimension is fixed to 128 and the penalty parameter of SVM is selected from {1e 3, 1e 2, 1e 1, 1, 1e2, 1e3}. |