Skip to content
Book Open access

CHARM-Mamba: Calibrated Hierarchical Adaptive Routing Multi-Source Mamba for Cross-City Traffic Prediction

Nov 2026 · Proceedings of the 35th ACM International Conference on Information and Knowledge Management · 0 citations · 41 references

Abstract

Cross-city transfer learning is pivotal for traffic prediction in data-scarce regions, yet impeded by structural limitations in existing frameworks. Conventional decoupled spatio-temporal modeling often struggles to fully capture the entangled, topology-dependent nature of congestion propagation. Moreover, indiscriminately aggregating heterogeneous source domains without granular pattern-specific filtering frequently induces negative transfer on incompatible data. Compounding these issues, full-parameter adaptation incurs prohibitive computational costs, severely hindering scalable few-shot deployment in dynamic real-world settings. To address these challenges, this paper proposes CHARM-Mamba, a Calibrated Hierarchical Adaptive Routing Multi-source transfer framework. First, the architecture integrates a Deep-Coupled Mamba backbone where State Space Models and dynamic spatial graphs mutually modulate to learn joint representations. Second, to mitigate negative transfer and support reliable knowledge reuse across heterogeneous sources, a routing mechanism on normalized feature spaces is introduced. By selectively retrieving compatible source prototypes via uncertainty-based calibration, it prioritizes transferable knowledge while discarding irrelevant noise. Third, to ensure parameter efficiency, a lightweight hypernetwork dynamically generates target-specific adapter parameters driven by routed patterns, facilitating pattern-aware adaptation with the backbone frozen. Extensive experiments demonstrate that CHARM-Mamba significantly outperforms state-of-the-art baselines in few-shot generalization. Notably, it achieves up to a 5.7% reduction in Mean Absolute Error (MAE) while utilizing fewer than 5% of the trainable parameters, highlighting its potential for scalable, trustworthy knowledge modeling in dynamic urban environments. The code and datasets used in this paper are available at: https://github.com/JOJOJISHENGWU/CHARM-Mamba.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.