Spatio-Temporal Attention-Fused Interactive Dynamic Graph Convolutional Network

2026-99-2440

11/20/2026

Authors
Abstract
Content
Traffic flow forecasting plays an important role in Intelligent Transportation Systems (ITS), and it directly helps with congestion control and resource planning in smart cities. Current graph-based deep learning methods have two main limitations in spatio-temporal modeling. These methods rely on fixed, manually set network topologies, which cannot reflect how real traffic changes over time. And they mostly use simple feature concatenation, which misses the deep coupling between spatial and temporal information. To solve these problems, this paper puts forward the Spatio-Temporal Attention-fused Interactive Dynamic Graph Convolutional Network (STA-IDGCN). The model uses a learnable dual-perspective memory bank to drive an adaptive dynamic graph generator. It does not let it follow physical road connections, and it rebuilds sparse graph structures through a Top-K adaptive filtering strategy. This helps uncover hidden spatial links while cutting down topological noise. To improve feature extraction, we use Coordinate Attention to highlight directional spatial data. An Efficient Channel Attention mechanism is also added inside an enhanced Gated Linear Unit. Cross-channel interactions can be managed and background noise can be blocked by this mechanism. An adaptive gating module is also designed to replace basic concatenation, with weights being assigned dynamically and historical time trends being tied with local spatial shifts at each prediction step. The full architecture combines several parts: firstly, periodic temporal embeddings; secondly, iterative diffusion spatial graph convolutions for tracing directional traffic; and thirdly, Transformer-based multi-head self-attention for long-range temporal patterns. The STA-IDGCN was evaluated on three widely used benchmark datasets: PEMS03, PEMS04, and PEMS08. Results show that our model gets better accuracy than existing baselines across all three datasets, confirming that interactive dynamic graph learning paired with long-range temporal modeling works effectively for traffic flow prediction.
Meta TagsDetails
DOI
https://doi.org/10.4271/2026-99-2440
Citation
Li, S. and He, Z., "Spatio-Temporal Attention-Fused Interactive Dynamic Graph Convolutional Network," 2026 2nd International Conference on Intelligent Transportation and Future Mobility (ITFM2026), Xi'An, China, March 27, 2026, https://doi.org/10.4271/2026-99-2440.
Additional Details
Publisher
Published
Nov 20
Product Code
2026-99-2440
Content Type
Technical Paper
Language
English