Research map / Architectures and efficiency / Attention and transformers
Satformer: Accurate and Robust Traffic Data Estimation for Satellite Networks
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 Satformer, an encoder-decoder tensor completion model with GCN graph embedding, adaptive sparse spatio-temporal attention (ASSIT), and a transfer module to estimate global inter-satellite traffic from sparse measurements on synthetic Iridium, Telesat and Starlink datasets.
Problem
Estimating global inter-satellite traffic matrices from partial, sparse measurements in large, highly dynamic satellite networks, where low-rank linear methods fail to capture nonlinear spatio-temporal correlations.
Approach
Combines GCN graph embedding with a sparse attention mechanism that masks attention to a central local window within regions, plus a self-attention transfer module between encoder and decoder, for tensor completion of satellite traffic.
Key results
- At 2% sampling on Iridium, Satformer reaches NMAE 0.098 vs. 0.1064 for the best neural baseline (STCAGCN), and 0.2782 for HaLRTC.
- On Starlink (1584 satellites) at 2% sampling, NMAE improves by 38.66% over the best baseline and NRMSE by 32.32%.
- Satformer reports the fastest training and inference times among compared methods, e.g. 80.3s training on Iridium vs. 164.6s for CoSTCo.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Satformer instead of SAITS | Satellite traffic estimation · Iridium, 2% sampling | NMAE: 0.1106 -> 0.098 | Helped |
| Use Satformer instead of SAITS | Satellite traffic estimation · Telesat, 2% sampling | NMAE: 0.1267 -> 0.1017 | Helped |
| Use Satformer instead of STCAGCN | Satellite traffic estimation · Starlink, 2% sampling | NMAE: 0.1944 -> 0.1402 | Helped |
| Add graph embedding (GE) (ablation) | Satformer · Satellite traffic datasets across varying sampling rates | Estimation performance | Helped |
| Add adaptive sparse spatio-temporal attention (ASSIT) (ablation) | Satformer · Satellite traffic datasets across varying sampling rates | Estimation performance | Helped |
| Add the transfer module (ablation) | Satformer · Satellite traffic datasets across varying sampling rates | Estimation performance | Helped |
Surprising findings
- Several baselines (e.g., LATC) show nearly flat error across sampling ratios, and some neural baselines show error increases at higher sampling ratios, suggesting unstable training.
Limitations
- Datasets are synthetically generated with a custom traffic generator because no public satellite traffic data exists, limiting realism validation.
- Interpretability is not addressed; the authors acknowledge deep models remain black boxes.
- Computational complexity still needs reduction for resource-limited satellites.
Experimental setup
Datasets: Iridium traffic (synthetic) Telesat traffic (synthetic) Starlink traffic (synthetic) Foursquare PeMS-Bay
Benchmarks: Iridium Telesat Starlink Foursquare PeMS-Bay
Metrics: NMAE NRMSE training time inference time
Baselines: HaLRTC LATC LETC CoSTCo CDSA DAIN SPIN SAITS STCAGCN
Builds on: GCN multi-head self-attention SAITS CoSTCo
Closest papers by meaning
- Decomposed Spatio-Temporal Mamba for Long-Term Traffic Prediction (AAAI 2025 · similarity 0.83)
- Efficient Traffic Prediction Through Spatio-Temporal Distillation (AAAI 2025 · similarity 0.81)
- Position: Mission Critical – Satellite Data is a Distinct Modality in Machine Learning (ICML 2024 · similarity 0.81)
- SSL-STMFormer Self-Supervised Learning Spatio-Temporal Entanglement Transformer for Traffic Flow Prediction (AAAI 2025 · similarity 0.81)
- TelTrans: Applying Multi-Type Telecom Data to Transportation Evaluation and Prediction via Multifaceted Graph Modeling (AAAI 2024 · similarity 0.81)
- Spatiotemporal-aware Trend-Seasonality Decomposition Network for Traffic Flow Forecasting (AAAI 2025 · similarity 0.80)
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.