STRAM: Spatio-temporal road-aware mapping for graph neural network prediction

Spatio-temporal graph neural networks (STGNNs) depend critically on how raw urban events are mapped to a graph. Yet, most work fixes the mapping to grids or administrative regions, which misalign with road structures and suffer from sparsity. We present STRAM, a road-aware, data-driven mapping method that (i) scores road nodes by a composite of network centrality, time-decayed event density, and spatial coverage, (ii) selects a resolution-controlled subset to define Voronoi regions, and (iii) builds edges that combine shared road length with a Gaussian distance kernel. STRAM is computed once, produces interpretable and application-tunable spatial partitions, and is compatible with any STGNN backbone. On NYC 311 service requests and motor-vehicle collision datasets, STRAM consistently improves STGNN models over fixed grids, administrative units, and degree-based Voronoi at matched graph sizes; for example, with DCRNN, it reaches 95.92% vs. 88.69% balanced accuracy on 311 and 85.58% vs. 78.11% on collision datasets. An ablation study further shows that the road network is the most influential component, while event density and spatial coverage exhibit dataset-dependent redundancy, underscoring the importance of adaptive weighting. STRAM is simple to deploy and provides robust improvements across architectures. © 2026 The Author(s)