STAAR: Spatio-Temporal Attention over Adaptive Regions for Traffic Forecasting

Abstract

Accurate traffic forecasting is essential for Intelligent Transportation Systems (ITS). Attention-based Spatio-Temporal Graph Neural Networks (STGNNs) have been widely adopted to capture long-range spatial dependencies, but their node-level attention incurs quadratic complexity and limits scalability in large-scale traffic networks. Although sparse and hierarchical attention mechanisms reduce this cost by restricting node pairs, they still rely on node-level modeling. Through empirical analysis on a 2,352-node benchmark, we find that removing node-level attention from representative baselines does not consistently degrade performance, suggesting that its additional predictive benefit may be marginal in large-scale settings. In contrast, region-level attention appears useful for modeling traffic propagation, which is also consistent with classical traffic engineering models that represent traffic dynamics over aggregated spatial units. Motivated by these observations, we propose STAAR, a Spatio-Temporal Attention over Adaptive Regions framework that dynamically constructs soft traffic regions and models dependencies among them. Specifically, adaptive region assignment assigns each node to multiple regions with normalized weights, while spatio-temporal attention captures time-lagged dependencies by flattening region tokens across time steps into a unified sequence. Extensive experiments on large-scale traffic benchmarks demonstrate that our model achieves competitive accuracy with substantial efficiency gains, reducing training time by 3.5×–29× and inference time by 2.8×–28× compared with existing attention-based STGNNs. The source code of STAAR is publicly available at https://github.com/kaist-dmlab/STAAR.