Traffic forecasting remains a fundamental yet challenging task in intelligent transportation systems, as it demands models capable of capturing complex and rapidly evolving spatiotemporal dynamics.
While existing approaches advanced the modeling of in...
Traffic forecasting remains a fundamental yet challenging task in intelligent transportation systems, as it demands models capable of capturing complex and rapidly evolving spatiotemporal dynamics.
While existing approaches advanced the modeling of inter-sensor relationships and addressed structural shifts through continual learning, they often incur high computational costs and struggle to respond effectively to sudden changes, limiting real-world applicability.
To overcome these issues, we introduce an online forecasting framework for traffic prediction and propose the Rank-one Adjacency Adapter (RoAA), a lightweight module designed for efficient, real-time edge-level adaptation.
RoAA is instantiated with two update mechanisms, exponential and softplus, yielding RoAA-Exp and RoAA-SP, respectively.
We further develop RoAA-Base, a compact traffic prediction model incorporating these adapters to support scalable online learning.
Empirical results on standard benchmark datasets demonstrate that RoAA-Base consistently outperforms existing models, achieving comparable or superior accuracy while requiring significantly fewer parameters and reduced online training time.
Comprehensive ablation studies and diagnostic analyses further confirm the contribution of each component to the overall performance.