Beyond Redundancy: Information-aware Unsupervised Multiplex Graph Structure Learning
Authors: Zhixiang Shen, Shuo Wang, Zhao Kang
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments against various baselines on different downstream tasks demonstrate its superior performance and robustness. Surprisingly, our unsupervised method even beats the sophisticated supervised approaches. |
| Researcher Affiliation | Academia | Zhixiang Shen , Shuo Wang , Zhao Kang School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu, Sichuan, China zhixiang.zxs@gmail.com zkang@uestc.edu.cn |
| Pseudocode | Yes | Algorithm 1: The optimization of Info MGF-RA; Algorithm 2: The optimization of Info MGF-LA |
| Open Source Code | Yes | The source code and datasets are available at https://github.com/zxlearningdeep/Info MGF. |
| Open Datasets | Yes | We conduct experiments on four real-world benchmark multiplex graph datasets, which consist of two citation networks (i.e., ACM [18] and DBLP [18]), one review network Yelp [35] and a large-scale citation network MAG [36]. |
| Dataset Splits | Yes | Table 5: Statistics of datasets. Columns for 'Training', 'Validation', 'Test' are provided with specific numbers for each dataset (e.g., ACM: 600 training, 300 validation, 2,125 test). |
| Hardware Specification | Yes | We implement all experiments on the platform with Py Torch 1.10.1 and DGL 0.9.1 using an Intel(R) Xeon(R) Platinum 8457C 20 v CPU and an L20 48GB GPU. |
| Software Dependencies | Yes | We implement all experiments on the platform with Py Torch 1.10.1 and DGL 0.9.1... |
| Experiment Setup | Yes | Our model is trained with the Adam optimizer, and Table 6 presents the hyper-parameter settings on all datasets. Here, E represents the number of epochs for training, and lr denotes the learning rate. The hidden-layer dimension dh and representation dimension d of graph encoder GCN are tuned from {32, 64, 128, 256}. The number of neighbors k for k NN is searched from {5, 10, 15, 20, 30}. The order of graph aggregation r and the number of layers L in GCN are set to 2 or 3... The probability ρ of random feature masking is set to 0.5 or 0, and the temperature parameter τc in contrastive loss is fixed at 0.2. For Info MGF-RA using random graph augmentation, the probability ρs of random edge dropping is fixed at 0.5. For Info MGF-LA with learnable generative graph augmentation, the generator s learning rate lrgen is fixed at 0.001, the temperature parameter τ in Gumbel-Max is set to 1, and the hyper-parameter λ controlling the minimization of mutual information is fine-tuned from {0.001, 0.01, 0.1, 1, 10}. |