Research map / Video / Video segmentation and tracking
Decomposed Spatio-Temporal Mamba for Long-Term Traffic Prediction
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 DST-Mamba, which decomposes traffic series into trend and seasonal parts, models seasonal parts with a bi-directional spatial Mamba over node tokens and trends with multi-scale linear predictors, achieving SOTA long-term traffic forecasting with near-linear cost.
Problem
Long-term traffic prediction, where entangled spatio-temporal dependencies are hard to disentangle and Transformer costs scale quadratically with sequence length.
Approach
Apply temporal decomposition to traffic series, then shift the seasonal part to a spatial (node-token) view processed by bi-directional Mamba with graph-aggregated node tokens and adaptive node embeddings, while trend parts go through multi-scale linear predictors with top-down scale mixing.
Key results
- On PEMS08 averaged over horizons, DST-Mamba reaches MSE 0.119 and MAE 0.209, versus 0.142/0.232 for S-Mamba's SOR-Mamba counterpart and 0.148/0.224 for S-Mamba.
- On Traffic (horizon 96-720 average), DST-Mamba achieves MSE 0.402 and MAE 0.270, matching the best MSE of S-Mamba/SOR-Mamba (0.402) and beating them on MAE.
- Ablation on PEMS08 shows removing decomposition raises average MSE from 0.119 to 0.139, and removing the seasonal component raises it to 0.513.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add temporal decomposition (ablation) | DST-Mamba · PEMS08 | MSE: 0.139 -> 0.119 | Helped |
| add the trend component (ablation) | DST-Mamba · PEMS08 | MSE: 0.134 -> 0.119 | Helped |
| add decomposition before the spatio-temporal Mamba module (ablation) | DST-Mamba · PEMS08 | MSE: 0.142 -> 0.119 | Helped |
| add adaptive node embeddings (ablation) | DST-Mamba · PEMS08 | MSE: 0.130 -> 0.119 | Helped |
| use bidirectional instead of unidirectional Mamba (ablation) | DST-Mamba · PEMS08 | MSE: 0.128 -> 0.119 | Helped |
| add the seasonal component (ablation) | DST-Mamba · PEMS08 | MSE: 0.513 -> 0.119 | Helped |
| add decomposition before the multi-scale linear prediction module (ablation) | DST-Mamba · PEMS08 | MSE: 0.551 -> 0.119 | Helped |
Surprising findings
- Removing the seasonal component causes a far larger degradation than removing the trend component, despite the trend part being motivated as the source of long-term patterns.
Limitations
- Computational cost is said to scale with node count rather than sequence length, so large-node graphs may still be costly.
- Hyperparameters such as trend weight, down-sampling window, and number of blocks are tuned on PEMS08 via sensitivity analysis.
Experimental setup
Datasets: Traffic PEMS03 PEMS04 PEMS07 PEMS08
Benchmarks: Traffic PEMS03 PEMS04 PEMS07 PEMS08
Metrics: MSE MAE
Baselines: iTransformer Crossformer PatchTST FEDformer Autoformer DLinear S-Mamba SOR-Mamba
Builds on: Mamba Autoformer FEDformer TimeMixer S-Mamba
Closest papers by meaning
- Spatiotemporal-aware Trend-Seasonality Decomposition Network for Traffic Flow Forecasting (AAAI 2025 · similarity 0.89)
- SSL-STMFormer Self-Supervised Learning Spatio-Temporal Entanglement Transformer for Traffic Flow Prediction (AAAI 2025 · similarity 0.89)
- Revisiting Spatial-Temporal Similarity: A Deep Learning Framework for Traffic Prediction (AAAI 2019 · similarity 0.89)
- Efficient Traffic Prediction Through Spatio-Temporal Distillation (AAAI 2025 · similarity 0.88)
- Towards Dynamic Spatial-Temporal Graph Learning: A Decoupled Perspective (AAAI 2024 · similarity 0.87)
- FlashST: A Simple and Universal Prompt-Tuning Framework for Traffic Prediction (ICML 2024 · similarity 0.86)
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.