A Sparse Cross Attention-Based Graph Convolution Network With Auxiliary Information Awareness for Traffic Flow Prediction

Abstract

Deep graph convolutional networks (GCNs) have shown promising performance in traffic prediction tasks, but their practical deployment on resource-constrained devices faces challenges. First, few models consider the potential influence of historical and future auxiliary information, such as weather and holidays, on complex traffic patterns. Second, the computational complexity of dynamic graph convolution operations grows quadratically with the number of traffic nodes, limiting model scalability. To address these challenges, this study proposes a deep encoder-decoder model named AIMSAN, which comprises an auxiliary information-aware module (AIM) and a sparse cross-attention-based graph convolutional network (SAN). From historical or future perspectives, AIM prunes multi-attribute auxiliary data into diverse time frames, and embeds them into one tensor. SAN employs a cross-attention mechanism to merge traffic data with historical embedded data in each encoder layer, forming dynamic adjacency matrices. Subsequently, it applies diffusion GCN to capture rich spatial-temporal dynamics from the traffic data. Additionally, AIMSAN utilizes the spatial sparsity of traffic nodes as a mask to mitigate the quadratic computational complexity of SAN, thereby improving overall computational efficiency. In the decoder layer, future embedded data are fused with feed-forward traffic data to generate prediction results. Experimental evaluations on three public traffic datasets demonstrate that AIMSAN achieves competitive performance compared to state-of-the-art algorithms, while reducing GPU memory consumption by 41.24%, training time by 62.09%, and validation time by 65.17% on average.