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..
HYPERION: Fine-Grained Hypersphere Alignment for Robust Federated Graph Learning
Authors: Frank Wan, Xiaoran Shang, Yuxin Wu, Guibin Zhang, Jinhe Bi, Liangtao Zheng, Xin Lin, Yue Liu, Yanbiao Ma, Wenke Huang, Bo Du
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we comprehensively evaluate HYPERION through four key axes: Q1 (Superiority), Q2 (Resilience), Q3 (Effectiveness), and Q4 (Sensitivity). The perfmance results are presented in Tab. 1, and Figure 3. |
| Researcher Affiliation | Academia | Guancheng Wan1 , Xiaoran Shang1 , Yuxin Wu4 , Guibin Zhang2, Jinhe Bi3, Liangtao Zheng5 , Xin Lin5, Yue Liu2, Yanbiao Ma4, Wenke Huang1 , Bo Du1 1Wuhan University 2NUS 3Ludwig-Maximilians-Universität München 4Renmin University of China 5UCSD EMAIL |
| Pseudocode | Yes | Algorithm 1 HYPERION Framework Communication rounds T, participant scale M, m-th client private model θm, m-th client local data Gm, m-th client prototype cluster Pm and loss weight α, β The final global model θglobal for t = 1, 2, , T do Client Side: for m = 1 to M in parallel do LCNC Hyperspherical Noise Calibration(Gm, Pm)by Equation (11) // Select robust nodes and train with them Sm Calculate Similarity(Pm)by Equation (6) // Calculate prototypes similarity metrix LST R Struct Loss(Sm)by Equation (8) // Calculate loss with interand intra-class prototypes LSEM Semantic Loss(Pm)by Equation (9) // Calculate loss with embedding vector and prototypes θt+1 m Local Updating(θt m, LCNC + αLST R + βLSEM) // Backward propagation Server Side: θpos, θneg GMM(P) // Client classification by prototypes Gneg m = Neighborhood Sparsification(θneg m , Gm), m by Equation (18) // pruning with Mahalanobis distance θglobal, Pglobal Aggregate(θpos m , Ppos), m // Clean clients hyperspherical aggregation P global Server Distillation(Pneg m , Pglobal), m by Equation (13) // Wassertein distance Server Distillation θm θglobal, m // Distribute parameters to clients return θglobal |
| Open Source Code | Yes | The code is available at: https://github.com/Guancheng Wan/HYPERION. |
| Open Datasets | Yes | Datasets. To effectively evaluate the performance of our approach, we utilize five benchmark graph datasets of various scales and distributions with different characteristics, including Cora [36], Cite Seer [11], Pub Med [3], Physics[43], and Amazon_ratings. |
| Dataset Splits | Yes | Each dataset is split into training, validation, and test sets in a fixed 20%/40%/40% ratio. |
| Hardware Specification | Yes | The experiments are conducted using NVIDIA Ge Force RTX 4090 GPUs as the hardware platform, coupled with Intel(R) Xeon(R) Platinum 8336C CPU @ 2.30GHz. |
| Software Dependencies | Yes | The deep learning framework employed was Pytorch, version 2.5.1, alongside CUDA version 12.2. |
| Experiment Setup | Yes | Our network features a fourlayer GCN backbone with uniform 384-dimensional hidden layers throughout the first three layers, each employing symmetric normalization (normalize=True) and Re LU activation, followed by 0.2 dropout for regularization. The final GCN layer produces compact 32-dimensional graph embeddings without activation. These embeddings are processed through a two-layer MLP head with Re LU activation in the hidden layer. The architecture optionally incorporates prototype learning with configurable parameters: each class maintains multiple 32-dimensional prototype vectors, and the prototype contrastive loss operates with a temperature parameter τ = 0.07 to control the similarity distribution sharpness. All GCN layers implement symmetric normalization (normalize=True), and consistent dropout (p = 0.2) is applied after each intermediate layer to prevent overfitting. TP-HSL parameter α is set in the range {0.40, 0.50, 0.60}, β Is set in the range {0.65, 0.70, 0.75}.As for GA-SHP parameter λ and η, we set λ in the range {0.03, 0.04}, η in the range {0.92, 0.94, 0.96}. The number of communication rounds is 100 for all methods. The number of clients M is set to 10 throughout all experiments, except for Figure 3 (Third). |