Preserving Microscopic State Integrity for Adaptive Modeling of Complex Spatio-Temporal Systems

Modeling the collective dynamics of complex systems, represented as graph-structured time series, is a central challenge in statistical physics and applied science. Prevailing Spatio-Temporal Graph Neural Networks (STGNNs) are constrained by a serial processing paradigm that creates structural information bottlenecks, leading to the premature abstraction and loss of fine-grained correlations between microscopic system states. This paper posits a hierarchical design principle: the efficacy of adaptive mechanisms, such as attention, is fundamentally contingent upon the integrity of the underlying feature propagation architecture. To validate this, we introduce DMA-EISTGCN, a framework featuring two innovations: (1) an Early Spatio-Temporal Interaction (EI) module, a non-serial design that ensures lossless feature fusion at the model’s front-end, and (2) a Dynamic Multi-scale Attention (DMA) mechanism that adaptively arbitrates between channel, spatial, and regional feature refinement. Experiments on four real-world traffic forecasting benchmarks show the model establishes a new state-of-the-art, reducing prediction error by up to 14.0 % compared to strong SOTA baselines. Critically, an ablation study reveals that applying the adaptive attention mechanism to a conventional serial backbone degrades performance. This provides direct empirical validation for our central thesis. This work establishes a new design paradigm for spatio-temporal models, demonstrating that resolving architectural bottlenecks is a necessary precursor to unlocking the full potential of adaptive learning in complex forecasting tasks.