Distribution and Aggregation: Continual Learning Principles for Evolving Traffic Networks

Real-world spatiotemporal networks continuously evolve, with new nodes emerging, traffic patterns changing, and network topology expanding. This evolution challenges spatiotemporal graph neural networks (STGNNs) to achieve continual learning: adapting to new data while mitigating catastrophic forgetting, scaling to growing networks without retraining from scratch, and maintaining prediction accuracy without incurring unbounded computational costs. The core dilemma is balancing model plasticity with stability as networks expand over time. To address these challenges, we propose DAA, a novel continual learning approach with multi-level adaptive tuning, following two fundamental principles guided by systematic analysis: distribution and aggregation, which effectively addresses the above issues through tensorized hierarchical parameters. Specifically, we deploy lightweight adapters with learnable importance weights across multiple network levels, leverage tensor-train decomposition to compress adapter representations, and jointly optimize with the base STGNN. This approach ensures that the model can sequentially learn from spatiotemporal data streams while adapting to network evolution and retaining learned knowledge. Experiments on multi-year traffic datasets demonstrate that DAA outperforms state-of-the-art baseline methods in effectiveness, generality, and efficiency. The code will be released at https://github.com/OvOYu/DAA.