Mamba-based model tames noisy urban traffic prediction with frequency-domain filtering

Traffic flow prediction models that work well on orderly highway networks often stumble when moved to city streets, where sensor readings are riddled with short-term fluctuations, missing local patterns and noise from complex urban traffic conditions. A research team including Xing Xu, Xu Zhao and Yun Zhao now reports a possible way around this fragility. In a regular paper published on 25 October 2026 in Volume 22 of the International Journal of Data Science and Analytics, the researchers introduce RoST-Mamba, a Robust Spatio-Temporal Mamba model that combines adaptive graph learning, frequency-domain enhancement and graph-constrained state-space modeling into a single forecasting architecture.

The core problem the authors identify lies in how existing spatiotemporal models handle structure and time. Most rely on fixed graph propagation or attention-based temporal modeling, which the paper argues leaves them poorly equipped when traffic signals contain strong high-frequency interference or when spatial dependencies between sensors shift over time. RoST-Mamba attacks both weaknesses directly.

It uses node-time embeddings to learn dynamic spatial correlations between sensors, so the graph that connects observation points is not frozen but adapts as conditions change. In parallel, a Top-K FFT-based frequency-domain module selects and amplifies the dominant periodic components of the signal while suppressing noisy spectral fluctuations. The third element ties space to time.

RoST-Mamba injects graph structure information into the state update process of its state-space model, meaning that temporal evolution is constrained by the learned spatial dependencies rather than unfolding independently of them. According to the authors, this graph-constrained state-space design is what allows the model to remain stable when urban sensor data degrade, because the network's learned geometry continually disciplines the temporal state as it advances through the forecast horizon. Experiments reported in the paper cover both open-source highway traffic datasets and a self-collected urban traffic dataset, allowing a direct test of the migration problem the model was designed to address.

RoST-Mamba achieved competitive prediction accuracy while maintaining low inference costs, a consideration for deployment in intelligent transportation systems where computational budgets are limited. On the urban transportation dataset, the model reached a mean absolute error of 5.21, a root mean square error of 7.96, and a mean absolute percentage error of 28.34 percent, figures the authors present as evidence that the architecture holds up under the noisier conditions of city road networks. Beyond the accuracy numbers, the paper includes frequency-domain visualization and ablation studies that the authors say verify the effectiveness of the two central design choices: the frequency-domain enhancement module and the graph-constrained modeling scheme.

Ablation studies remove components of a model one at a time to measure each part's contribution, and here they were used to show that performance depends on the pieces RoST-Mamba adds to the standard spatiotemporal toolkit. The work, whose author list also includes Guoming Qian, Na Wu, Yali Wang, Petr Skobelev and Yanan Mi, positions state-space models as an alternative to attention-based approaches for traffic forecasting, particularly where robustness on noisy urban data is the binding constraint.

Subject of Research: Machine learning model for robust traffic flow prediction on urban road networks

Article Title: RoST-Mamba: traffic prediction model based on adaptive graph and frequency-domain enhancement

Article References: Xu, X., Zhao, X., Zhao, Y., Qian, G., Wu, N., Wang, Y., Skobelev, P., & Mi, Y. (2026). RoST-Mamba: traffic prediction model based on adaptive graph and frequency-domain enhancement. International Journal of Data Science and Analytics, 22(1), 357. https://doi.org/10.1007/s41060-026-01281-9

Image Credits: AI Generated

DOI: 10.1007/s41060-026-01281-9

Keywords: traffic prediction, RoST-Mamba, state-space models, adaptive graph learning, frequency-domain enhancement, intelligent transportation systems, spatiotemporal modeling, urban road networks, Mamba, machine learning, traffic, prediction