Research map / Federated learning and privacy / Personalized and heterogeneous FL
Personalized Federated Learning for Spatio-Temporal Forecasting: A Dual Semantic Alignment-Based Contrastive Approach
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 FUELS, a personalized federated learning method for spatio-temporal (wireless traffic) forecasting that uses dual contrastive tasks, a learnable hard-negative filtering module for intra-client temporal contrast, and JSD-based client-customized global prototypes for inter-client spatial contrast.
Problem
Federated spatio-temporal forecasting fails to capture spatio-temporal heterogeneity across clients; existing contrastive methods cannot reliably determine positive/negative pairs and do not fit the personalized FL paradigm without exposing raw data or sending large parameters.
Approach
Align positive and negative pairs by semantic (cosine) similarity via a learnable hard-negative filtering matrix for an intra-client contrastive loss, and use a periodicity-aware averaged prototype communicated to the server, where JSD-based similarity yields client-customized global positive/negative prototypes for an inter-client contrastive loss.
Key results
- FUELS achieves average MSE 0.827 and MAE 0.446 across SMS, Call and Net, versus 0.917 MSE for the best baseline PerFedAvg.
- Communication parameters per client per round drop from 100,737 (FedAvg and other baselines) to 6,144 for FUELS, a reduction of about 94%.
- On the Net dataset FUELS reaches MSE 0.880 versus 1.107 for PerFedAvg.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FUELS instead of pFedMe | Federated spatio-temporal forecasting · SMS | MSE: 1.250 -> 1.249 | Helped |
| Use FUELS instead of pFedMe | Federated spatio-temporal forecasting · SMS | MAE: 0.549 -> 0.541 | Helped |
| Use FUELS instead of FedRep | Federated spatio-temporal forecasting · Call | MSE: 0.372 -> 0.353 | Helped |
| Use FUELS instead of FedAvg | Federated spatio-temporal forecasting · Call | MAE: 0.300 -> 0.311 | Hurt |
| Use FUELS instead of PerFedAvg | Federated spatio-temporal forecasting · Net | MSE: 1.107 -> 0.880 | Helped |
| Use FUELS instead of FedRep | Federated spatio-temporal forecasting · Net | MAE: 0.625 -> 0.488 | Helped |
| Add the inter-client contrastive loss (ablation) | FUELS · Net | prediction error | Helped |
| Add the intra-client contrastive loss (ablation) | FUELS · Net | prediction error | Helped |
| Use periodicity-aware prototypes instead of concatenation-based prototypes (ablation) | FUELS · Net | prediction error | Helped |
| Add dynamic hard negative filtering (ablation) | FUELS · Net | prediction error | Helped |
Surprising findings
- Several baselines (e.g., FedDA) underperform the Solo local-only training on average MSE.
Limitations
- Relies on a periodicity assumption (batch size set equal to traffic periodicity) for the periodicity-aware prototype.
- JSD pairwise computation scales as O(N^2), mitigated only by random client sampling.
- Theoretical guarantees rest on assumptions such as bounded loss and Lipschitz continuity that are not verified empirically.
Experimental setup
Datasets: SMS Call Net
Benchmarks: SMS Call Net
Metrics: MSE MAE communication parameters per round
Baselines: Solo FedAvg FedProx FedRep PerFedAvg pFedMe FedDA
Builds on: FedAvg FedRep pFedMe contrastive learning
Closest papers by meaning
- Time-FFM: Towards LM-Empowered Federated Foundation Model for Time Series Forecasting (NeurIPS 2024 · similarity 0.85)
- FedTGP: Trainable Global Prototypes with Adaptive-Margin-Enhanced Contrastive Learning for Data and Model Heterogeneity in Federated Learning (AAAI 2024 · similarity 0.85)
- Relaxed Contrastive Learning for Federated Learning (CVPR 2024 · similarity 0.85)
- ST-ReP: Learning Predictive Representations Efficiently for Spatial-Temporal Forecasting (AAAI 2025 · similarity 0.85)
- Handling Spatial-Temporal Data Heterogeneity for Federated Continual Learning via Tail Anchor (CVPR 2025 · similarity 0.84)
- FedGPS: Statistical Rectification Against Data Heterogeneity in Federated Learning (NeurIPS 2025 · similarity 0.83)
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.