Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
OASIS: One-Shot Federated Graph Learning via Wasserstein Assisted Knowledge Integration
Authors: Frank Wan, Jiaru Qian, Wenke Huang, Qilin Xu, Xianda Guo, Boheng Li, Guibin Zhang, Bo Du, Mang Ye
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on real-world tasks demonstrate the superior performance and generalization capability of OASIS, with an average improvement of 15.81% over the baseline. The code is available for anonymous access at https://github.com/Jiaru Qian/OASIS. |
| Researcher Affiliation | Academia | Guancheng Wan1 , Jiaru Qian1 , Wenke Huang1 , Qilin Xu1, Xianda Guo1, Boheng Li2, Guibin Zhang3, Bo Du1, Mang Ye1 1School of Computer Science, Wuhan University 2NTU 3NUS EMAIL |
| Pseudocode | No | The paper does not contain clearly labeled pseudocode or algorithm blocks. The methods are described in narrative text and mathematical formulations. |
| Open Source Code | Yes | The code is available for anonymous access at https://github.com/Jiaru Qian/OASIS. |
| Open Datasets | Yes | Datasets. To effectively evaluate the performance of our approach, we employed eight benchmark graph datasets of various scales and distributions, including Cora [37], Cite Seer [10], Pub Med [5], Amazon-Photo, Coauthor-CS [45], Actor [42], Roman-empire [43] and Ogbn-Arxiv [17]. Detailed descriptions and splits for these datasets can be found in Appendix D. |
| Dataset Splits | Yes | Each dataset is split into training, validation, and test sets in a fixed 20%/40%/40% ratio. The key statistics of these datasets are summarized in Tab. 3. |
| Hardware Specification | Yes | The experiments are conducted using NVIDIA Ge Force RTX 3090 GPUs as the hardware platform, coupled with Intel(R) Xeon(R) Gold 6240 CPU @ 2.60GHz. |
| Software Dependencies | Yes | The deep learning framework employed was Pytorch, version 2.3.1, alongside CUDA version 12.1. |
| Experiment Setup | Yes | Implementation Details. We adopt a two-layer GCN as the backbone, with a hidden layer size of 128. We set K = 10 clients and draw pk Dir(α) from a Dirichlet distribution [40] and assign a fraction pc k of class c to client k. Specifically, α is set as 0.05 to simulate a highly non-IID senario. The codebook size is set in the range {26, 27, 28}. More implementation details and experiments on various client numbers can be found in Appendix F and Appendix G. As for optimization of the graph synthesizer, Adaptive Moment Estimation (Adam) was chosen, featuring a learning rate of 5e 3 and a weight decay of 4e 4. The codebook size is set in the range {26, 27, 28}, with the same optimizer and learning parameter. At the local training phase, we set the training epoch TS of the synthesizer to 100 and epoch TC of the teacher GNN and the codebook to 50. λd and λf are determined through a grid search [32] within {0.01, 0.05, 0.1, 0.5} and {0.1, 0.2, 0.5, 1} respectively. η, λo are set as 0.25, 0.01 and λc is set to 1. To make sure that LFGW is on the same scale as other loss functions for Amz-Photo and Ogbn-Arxiv datasets, we set their λf scales to 1e 5 and 1e 7, respectively. We set a in LFGW as 0.5 to balance the feature part and the structure part. The communication round is limited to one. At the server side, we determine the global distilling epoch in the range {10, 20, 30} and adopt Adam as the optimizer for the global model with a learning rate of 1e 2 and a weight decay of 4e 4. The synthesized graphs generated by the synthesizer of each client have the same scale ˆN k as the corresponding local subgraph. For the large graph Ogbn-Arxiv, we set TC to 1 and ˆN k to one-tenth of the local graph. Moreover, λk, κ is set to 0.01, 1 and λs is determined in range {1, 5, 10}. The distillation temperature τ is set to 3 for all datasets. |