Federated Node-Level Clustering Network with Cross-Subgraph Link Mending

Authors: Jingxin Liu, Renda Han, Wenxuan Tu, Haotian Wang, Junlong Wu, Jieren Cheng

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate the superiority of Fed NCN. Extensive experiments on five graph benchmark datasets demonstrate the effectiveness and superiority of the proposed Fed NCN compared to its competitors.
Researcher Affiliation Academia 1School of Cyberspace Security, Hainan University, Haikou, China 2School of Computer Science and Technology, Hainan University, Haikou, China. Correspondence to: Wenxuan Tu <EMAIL>, Jieren Cheng <EMAIL>.
Pseudocode Yes Algorithm 1 Training Procedure of Fed NCN Algorithm 2 Fed NCN for Client Algorithm Algorithm 3 Fed NCN for Server Algorithm
Open Source Code No The paper does not provide concrete access to source code for the methodology described.
Open Datasets Yes Specifically, we use Cite Seer (Liu et al., 2023a), Pub Med (Jiang et al., 2024), Amazon-Computer, Amazon-Photo (Lin et al., 2021), and Questions (Platonov et al., 2024) as our experimental benchmark datasets.
Dataset Splits Yes Following the experimental setup from Fed TAD (Zhu et al., 2024), we construct distributed subgraphs by dividing the dataset into 5 clients, 10 clients, and 20 clients, respectively, where each client has a subgraph that is part of a complete graph.
Hardware Specification Yes All methods are implemented using Py Torch 2.4.0 and a single NVIDIA Ge Force RTX 4090 GPU.
Software Dependencies Yes All methods are implemented using Py Torch 2.4.0 and a single NVIDIA Ge Force RTX 4090 GPU.
Experiment Setup Yes We utilize a four-layer GNN on both the client and the server to obtain node embeddings, with hidden layer dimensions are 500-500-2000-10. Moreover, we use a one-layer MLP to obtain the local clustering signals, which are then uploaded to the server. During model optimization, we adopt the Adam optimizer (Xiao et al., 2024) with a learning rate of 1e-3. The client-server interaction is conducted 20 times, with the local model training 10 epochs during each interaction.