Spiking Graph Neural Network on Riemannian Manifolds
Authors: Li Sun, Zhenhao Huang, Qiqi Wan, Hao Peng, Philip S Yu
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on common graphs show the proposed MSG achieves superior performance to previous spiking GNNs and energy efficiency to conventional GNNs. |
| Researcher Affiliation | Academia | Li Sun North China Electric Power University Beijing 102206, China ccesunli@ncepu.edu.cn Zhenhao Huang North China Electric Power University Beijing 102206, China huangzhenhao@necpu.edu.cn Qiqi Wan North China Electric Power University Beijing 102206, China wanqiqi@ncepu.edu.cn Hao Peng Beihang University Beijing 100191, China penghao@buaa.edu.cn Philip S. Yu University of Illinois at Chicago IL, USA psyu@uic.edu |
| Pseudocode | Yes | Algorithm 1 Training MSG by the proposed Differentiation via Manifold |
| Open Source Code | Yes | 2Codes are available at https://github.com/Zhenh Huang/MSG |
| Open Datasets | Yes | Our experiments are conducted on 4 commonly used benchmark datasets including two popular co-purchase graphs: Computers and Photo[53], and two co-author graphs: CS and Physics [53]. Datasets are detailed in Appendix E. |
| Dataset Splits | No | The paper refers to common benchmark datasets and evaluation protocols but does not explicitly state the training, validation, and test splits (e.g., percentages or sample counts) within its own text. |
| Hardware Specification | Yes | Experiments are conducted on the hardware of NVIDIA Ge Force RTX 4090 GPU 24GB memory, and AMD EPYC 9654 CPU with 96-Core Processor. |
| Software Dependencies | No | Our model is built upon Geo Opt [56], Spiking Jelly [56] and Py Torch [57]. |
| Experiment Setup | Yes | The dimension of the representation space is set as 32. The manifold spiking neuron is based on the IF model [49] by default... The time steps T for neurons is set to 5 or 15. The step size ϵ in Eq. 8 is set to 0.1. The hyperparameters are tuned with grid search, in which the learning rate is {0.01, 0.003} for node classification and {0.003, 0.001} for link prediction, and the dropout rate is in {0.1, 0.3, 0.5}. |