Bi-Directional Multi-Scale Graph Dataset Condensation via Information Bottleneck
Authors: Xingcheng Fu, Yisen Gao, Beining Yang, Yuxuan Wu, Haodong Qian, Qingyun Sun, Xianxian Li
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Encouraging empirical results on several datasets demonstrates the significant superiority of the proposed framework in graph condensation at different scales. |
| Researcher Affiliation | Academia | 1Guangxi Key Lab of Multi-source Information Mining & Security, Guangxi Normal University, Guilin, China 2Key Lab of Education Blockchain and Intelligent Technology, Ministry of Education, Guangxi Normal University, China 3Institute of Artificial Intelligence, Beihang University, Beijing, China 4School of Computer Science and Engineering, Beihang University, Beijing, China 5University of Edinburgh, Edinburgh, UK |
| Pseudocode | Yes | Algorithm 1: Bi-directional Condensation Algorithm |
| Open Source Code | Yes | To evaluate the performance of our Bi MSGC 1, we choose five node classification benchmark graphs, including three transductive graphs, Cora, Citeseer (Kipf and Welling 2017), Ogbn-Arxiv (Hu et al. 2021a) and two inductive graphs, Flickr and Reddit (Zeng et al. 2020). (footnote 1: https://github.com/Ring BDStack/Bi MSGC.) |
| Open Datasets | Yes | Datasets. To evaluate the performance of our Bi MSGC 1, we choose five node classification benchmark graphs, including three transductive graphs, Cora, Citeseer (Kipf and Welling 2017), Ogbn-Arxiv (Hu et al. 2021a) and two inductive graphs, Flickr and Reddit (Zeng et al. 2020). |
| Dataset Splits | No | For the needs of the multi-scale graph dataset condensation task, we distilled all graphs to the largest scale and then tested them by random sampling the subgraph according to the target reduction rate. |
| Hardware Specification | Yes | All models were trained and tested on a single Nvidia A800 80GB GPU. |
| Software Dependencies | No | Baselines. Based on GC-Bench (Sun et al. 2024) 2, we select gradients-based methods: GCond (Jin et al. 2022), SGDD (Yang et al. 2023), EXGC (Fang et al. 2024); trajectory-based methods: SFGC (Zheng et al. 2023),GEOM (Zhang et al. 2024); other method: GCDM (Liu et al. 2022), GC-SNTK (Liu et al. 2022), GDEM (Liu, Bo, and Shi 2024). (footnote 2: https://github.com/Ring BDStack/GC-Bench.) |
| Experiment Setup | Yes | Algorithm 1: Bi-directional Condensation Algorithm Input: Original Graph G = {X, A}; Number of training meso-scale epochs E1; Number of training bi-directional scale epochs E2, the learning rate η. [...] Settings and Hyperparameters. We followed the default values of parameters for baselines. [...] The remaining details are given in the Appendix. |