Research map / Federated learning and privacy / Personalized and heterogeneous FL
PICSR: Prototype-Informed Cross-Silo Router for Federated Learning (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PICSR, a federated mixture-of-experts whose router is grounded in silo prototypes (mean samples), so routing weights can explain inter-institution differences on heart disease prediction.
Problem
Federated learning mostly optimizes predictive accuracy and yields black-box models, offering little explanation of how institutions differ and what they gain from collaboration in cross-silo settings.
Approach
A mixture-of-experts over silo-specific local models whose router computes inverted L2 distances between embedded samples and embedded silo prototypes (per-silo mean samples), so router weights are grounded in each silo's data distribution and can be inspected by stakeholders.
Key results
- On FHD, PICSR reports 77.88% on Cleveland, 80.67% on Hungary, 90.00% on Switzerland and 82.67% on VA (mean over 5 runs), outperforming an ensemble on most hospitals and most federated baselines.
- Routing distributions show Switzerland's test samples mostly routed to its own model, while Cleveland and Hungary share routed samples, reflecting similar distributions and benefits of collaboration.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use prototype-informed routing to weight local experts instead of averaging their predictions | FHD · Cleveland | accuracy: 79.81±1.92 -> 77.88±2.11 | Hurt |
| Use prototype-informed routing to weight local experts instead of averaging their predictions | FHD · Hungary | accuracy: 78.88±2.29 -> 80.67±1.31 | Helped |
| Use prototype-informed routing to weight local experts instead of averaging their predictions | FHD · Switzerland | accuracy: 81.25±3.95 -> 90.00±3.06 | Helped |
| Use prototype-informed routing to weight local experts instead of averaging their predictions | FHD · VA | accuracy: 77.78±3.14 -> 82.67±1.66 | Helped |
Limitations
- Evaluated on a single small tabular heart disease dataset with four silos
- Prototypes are simple feature means, which may be a crude summary of silo distributions for other data types
- Explanations rely on qualitative inspection of routing plots rather than a quantitative evaluation
Experimental setup
Datasets: Fed-Heart-Disease (FHD)
Benchmarks: Fed-Heart-Disease
Metrics: accuracy routing distribution
Baselines: Local Ensemble FedAvg FedAvg + Fine-tuning FedProx
Builds on: FedAvg FedProx mixture of experts
Closest papers by meaning
- FedCompass: Efficient Cross-Silo Federated Learning on Heterogeneous Client Devices Using a Computing Power-Aware Scheduler (ICLR 2024 · similarity 0.84)
- Local Superior Soups: A Catalyst for Model Merging in Cross-Silo Federated Learning (NeurIPS 2024 · similarity 0.84)
- Cross-Silo Feature Space Alignment for Federated Learning on Clients with Imbalanced Data (AAAI 2025 · similarity 0.84)
- Taming Cross-Domain Representation Variance in Federated Prototype Learning with Heterogeneous Data Domains (NeurIPS 2024 · similarity 0.83)
- Free-Rider and Conflict Aware Collaboration Formation for Cross-Silo Federated Learning (NeurIPS 2024 · similarity 0.82)
- FedGPS: Statistical Rectification Against Data Heterogeneity in Federated Learning (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.