ML$^2$-GCL: Manifold Learning Inspired Lightweight Graph Contrastive Learning
Authors: Jianqing Liang, Zhiqiang Li, Xinkai Wei, Yuan Liu, Zhiqiang Wang
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive empirical results on various benchmarks and evaluation protocols demonstrate effectiveness and lightweight of ML2-GCL. We conduct a series of experiments to demonstrate the superiority of ML2-GCL. First, we briefly describe the datasets. Then, we evaluate the empirical performance across various graph datasets on node classification and link prediction tasks. In the end, we present the ablation study, hyperparameters analysis and embeddings visualization results. |
| Researcher Affiliation | Academia | 1Key Laboratory of Computational Intelligence and Chinese Information Processing of Ministry of Education, School of Computer and Information Technology, Shanxi University, Taiyuan 030006, Shanxi, China. Correspondence to: Jianqing Liang <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 ML2-GCL Input: Graph G = (V, E) Output: Node embeddings Z |
| Open Source Code | Yes | We release the code at https://github.com/a-hou/ML2-GCL. |
| Open Datasets | Yes | We conduct experiments on 6 widely-used datasets including Cora, Citeseer, Pubmed, Amazon-Photo, Amazon Computers and Wiki-CS. The detailed statistics are summarized in Table 4. Cora, Citeseer and Pub Med (Sen et al., 2008) are well-known citation network datasets... Amazon-Photo and Amazon-Computers (Shchur et al., 2018) are two networks... Wiki-CS (Mernyei & Cangea, 2020) is a reference network... |
| Dataset Splits | Yes | For node classification task, we split Cora, Citeseer and Pubmed for the training, validation and testing following (Yang et al., 2016), and all the other datasets following (Liu et al., 2020). For link prediction task, we follow the experimental setup of GCA (Zhu et al., 2021). For each dataset, we conduct 20 random splits of training/validation/test, and report the average performance of all algorithms on the same random splits. |
| Hardware Specification | Yes | We implement all experiments on the platform with Py Torch 2.0.1 and Py Torch Geometric 2.6.1 on NVIDIA 3090 GPUs with 24GB memory. |
| Software Dependencies | Yes | We implement all experiments on the platform with Py Torch 2.0.1 and Py Torch Geometric 2.6.1 on NVIDIA 3090 GPUs with 24GB memory. |
| Experiment Setup | Yes | For chosen hyper-parameters see Appendix C. In hyperparameter search, we attempt to adjust the value of k and λ in ML2-GCL, as well as other deep learning hyperparameters including temperature parameter τ, hidden dim, learning rate, dropout and weight decay. We apply the grid search strategy to choose the optimal hyperparameters. Specifically, we search k in [1, 5], λ in {0, 0.01, 0.1, 1, 10, 100}, hidden dim from {128, 256, 500, 512}, learning rate from {0.0005, 0.001, 0.005, 0.01} and weight decay in {5e-5, 1e-4, 5e-4, 7e-4}. Tables 5-6 give the hyperparameters specifications for ML2-GCL on node classification and link prediction tasks. |