Abstract: As a pivotal technique in traffic flow forecasting, Graph Convolutional Recurrent Networks (GCRN) effectively enhance prediction performance. However, given the pronounced spatio-temporal heterogeneity inherent in traffic data, existing GCRNs typically generate parameters independently for each node or time step, resulting in massive model parameter counts and high redundancy. To alleviate the aboveissues, an Adaptive Meta-node Graph Convolutional Recurrent Network (AMGCRN) was constructed by leveraging the advantages of meta-nodes in reducing parameter redundancy, as well as the capability of hypernetworks to adapt to spatio-temporal heterogeneity while balancing performance and efficiency. First, a small set of learnable meta-nodes was introduced to distill differentiated traffic patterns. Then, a hypernetwork was employed to dynamically generate node-specific convolutional parameters for different time steps, thereby allowing spatio-temporal heterogeneity to be accommodated while parameter redundancy was effectively reduced. In addition, a multi-head cross-attention mechanism was incorporated to enhance the capture of global temporal dependencies. Experimental results on multiple real-world traffic datasets demonstrate that, compared to Meta Attentive Graph Convolutional Recurrent Network (MAGCRN), the proposed model achievesa relative reduction of approximately 1%-3% in terms of MAE (Mean Absolute Error). Furthermore, in the efficiency evaluation on the PeMSD8 dataset, the proposed model reduces the number of parameters by 83% and FLOPs by 22% compared with MAGCRN, thereby realizing a comprehensive optimization across model capacity, computational complexity, and predictive performance.

Key words: traffic flow prediction, Graph Convolutional Recurrent Network (GCRN), meta-node, hypernetwork, attention mechanism

摘要： 图卷积循环网络(GCRN)作为交通流量预测中的关键技术之一，可有效提升交通流量预测性能，但由于交通数据在时空维度上表现出强烈的异质性，现有的GCRN通常为每个节点或时间步独立生成参数，导致模型参数庞大且冗余度高。为缓解上述问题，并且充分利用元节点减少参数冗余的优势，以及超网络适配时空异质性、平衡性能与效率的特性，构建基于自适应元节点学习的图卷积循环网络(Adaptive Meta-node Graph Convolutional Recurrent Network，AMGCRN)。该模型首先通过引入少量可学习的元节点以凝练差异化交通模式，并利用超网络动态生成各节点在各时间步的专属卷积参数，不仅可以保证时空异质性，而且有效避免参数冗余；其次，通过多头交叉注意力机制，确保对全局时间依赖的捕捉能力。在多个真实交通数据集上的实验结果表明，相较于元注意力图卷积循环网络(MAGCRN)，所提模型在平均绝对误差(MAE)指标上的相对降幅约为1%~3%。进一步地，在PeMSD8数据集上的效率评估中，所提模型的参数量相较于MAGCRN减少了83%，FLOPs降低了22%，实现了模型容量、计算复杂度与预测性能的综合优化。

关键词: 交通流量预测, 图卷积循环网络, 元节点, 超网络, 注意力机制