Skip to content

FAST: A Synergistic Framework of Attention and State-space Models for Spatiotemporal Traffic Prediction

Home / Papers / FAST: A Synergistic Framework of Attention and State-space Models for Spatiotemporal Traffic Prediction

FAST: A Synergistic Framework of Attention and State-space Models for Spatiotemporal Traffic Prediction

Authors: Xinjin Li, Jinghan Cao, Mengyue Wang, Yue Wu, Longxiang Yan, Yeyang Zhou, Ziqi Sha, Yu Ma Date: 2026-04-15 Paper ID: openalex:2604.13453

Summary

FAST is a unified spatiotemporal framework designed for scalable traffic forecasting by combining temporal attention with selective state-space modeling. The model utilizes a Temporal-Spatial-Temporal architecture to efficiently capture both complex temporal patterns and long-range spatial dependencies across large sensor networks. By integrating a Mamba-based spatial module, FAST maintains linear complexity while achieving superior predictive accuracy compared to existing Transformer and GNN-based baselines across standard traffic benchmarks.

Key Contributions

  • Introduces FAST, a Temporal-Spatial-Temporal architecture that synthesizes attention mechanisms with state-space models for scalable traffic forecasting.
  • Employs a Mamba-based spatial module to model inter-sensor dependencies with linear complexity, overcoming efficiency limitations of traditional Transformer architectures.
  • Achieves state-of-the-art performance on PeMS04, PeMS07, and PeMS08 datasets, outperforming existing Transformer, GNN, and Mamba-based approaches.

Open Questions & Future Work

Archivist Review

The paper introduces a hybrid architecture combining attention and state-space models. While effective, the individual components and the specific architecture are variations of existing paradigms. I have consolidated the suggested open question into a more precise formulation regarding adaptive spatial modeling and scale, as the original was too broad. Standard benchmarks like PeMS were rejected as they are routine in this domain.

Approved Open Questions

  • Adaptive Spatiotemporal Forecasting Scalability: This addresses the fundamental transition from controlled benchmark forecasting to reliable deployment in complex, evolving urban infrastructure.

Rejected Candidates

  • [dataset] PeMS datasets (pems04-pems07-pems08) - not_novel: These are standard, routine traffic forecasting benchmarks rather than unique or novel datasets.
  • [concept] Temporal-Spatial-Temporal architecture (temporal-spatial-temporal-architecture) - subcomponent_of_broader_mechanism: This is a specific architectural composition (sandwich structure) for this paper that is not sufficiently generalizable as a standalone concept.
  • [concept] Mamba-based spatial module (mamba-based-spatial-module) - subcomponent_of_broader_mechanism: This is a straightforward application of Mamba to a specific domain subcomponent rather than a novel conceptual contribution.

Metadata & Links

url
https://arxiv.org/abs/2604.13453
paper_id
2604.13453
paper_source
openalex
domain
time-series
tags
architectures
datasets
concept_slugs
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-04-18T05:35:06Z