FedGCN: Convergence-Communication Tradeoffs in Federated Training of Graph Convolutional Networks
Authors: Yuhang Yao, Weizhao Jin, Srivatsan Ravi, Carlee Joe-Wong
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show that our Fed GCN algorithm achieves better model accuracy with 51.7% faster convergence on average and at least 100 less communication compared to prior work1. Our experiments on both synthetic and real-world datasets demonstrate that Fed GCN outperforms existing distributed GCN training methods in most cases with a fast convergence rate, higher accuracy, and orders-of-magnitude lower communication cost. |
| Researcher Affiliation | Academia | Yuhang Yao Carnegie Mellon University yuhangya@andrew.cmu.edu Weizhao Jin University of Southern California weizhaoj@usc.edu Srivatsan Ravi University of Southern California sravi@isi.edu Carlee Joe-Wong Carnegie Mellon University cjoewong@andrew.cmu.edu |
| Pseudocode | No | Based on the insights in the previous section, we introduce the Fed GCN training algorithm (details are provided in Appendix A). The appendix content is not included in the provided paper text. |
| Open Source Code | Yes | 1Code in https://github.com/yh-yao/Fed GCN |
| Open Datasets | Yes | We use the Cora (2708 nodes, 5429 edges), Citeseer (3327 nodes, 4732 edges), Ogbn-Ar Xiv (169343 nodes, 1166243 edges), and Ogbn-Products (2449029 nodes, 61859140 edges) datasets to predict document labels and Amazon product categories (Wang et al., 2020b; Hu et al., 2020). |
| Dataset Splits | No | The paper describes using training nodes (Vtrain) and assigning labels to other nodes (V\Vtrain), but it does not explicitly provide specific percentages, sample counts, or a distinct 'validation' split for reproducibility. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory, or cloud instance types) used for running the experiments. |
| Software Dependencies | No | The paper does not provide specific software dependencies or library version numbers (e.g., Python 3.8, PyTorch 1.9) used in the experiments. |
| Experiment Setup | No | For Cora and Citeseer, we use a 2-layer GCN with Kipf and Welling (2016) s hyper-parameters. For Ogbn-Arxiv and Ogbn-Products, we respectively use a 3-layer GCN and a 2-layer Graph Sage with Hu et al. (2020) s hyper-parameters. Detailed experimental setups and extended results, including an evaluation of the HE overhead, are in Appendix E and G. |