HEPHAESTUS: Hierarchical Periodic Heterogeneous Adaptive Spatio-Temporal Unified System for Traffic Forecasting

ShubinRuan, Weixin Pang, SaiminYe, LijingYan, haoyonggang, Weihao Jiang

Keywords: traffic forecasting, spatio-temporal modeling, adaptive multi-scale, periodic temporal attention, heterogeneous spatial attention

Abstract: Accurate traffic forecasting requires modeling complex spatio-temporal dynamics characterized by multi-scale temporal patterns, periodic dependencies, and spatial heterogeneity. While recent advances in spatio-temporal graph neural networks (STGNNs) have improved predictive performance, they often rely on fixed architectures that lack adaptivity to input-driven variations in temporal granularity and spatial connectivity. In this work, we propose \textbf{HEPHASTUS}, a novel framework that unifies adaptive multi-scale temporal modeling, explicit periodicity-aware attention, and dynamic spatial heterogeneity learning within a lightweight, scalable architecture. Our approach introduces three key components: (i) an Adaptive Multi-Scale Mixture of Experts (AMS-MoE) that deeply integrates multi-scale modeling with expert routing, using a dynamic router to automatically assign input sequences of different time scales to specialized experts, each expert focuses on temporal feature extraction at a specific scale (e.g., local fluctuations or long-term trends), with scale weights adaptively adjusted according to the time-varying input characteristics, enabling collaborative capture of global dependencies and local details; (ii) a Periodic Temporal Attention (PTA) mechanism that explicitly captures daily and weekly patterns via parameterized period matrices; and (iii) a Heterogeneous Spatial Attention (HSA) module that balances global structure and local specificity through node embeddings and a learnable pattern library, with low parametric cost. Experiments on six real-world traffic datasets, METR-LA, PEMS-BAY, PEMS03 and others, show that the proposed method achieves state-of-the-art performance across MAE, RMSE, and MAPE, with consistent gains over existing baselines. Ablation studies confirm the necessity of each design choice. Our results highlight the importance of adaptive, structured modeling in capturing the intrinsic dynamics of urban traffic.

Primary Area: learning on time series and dynamical systems

Submission Number: 10198