LCA-STGNet: A Multimodal Dynamic Adjacency Spatiotemporal Graph Neural Network for Traffic Prediction

In recent years, with the in-depth application of AI technology in the field of spatiotemporal fusion, modeling of complex spatiotemporal dependencies of small sample data has become a hot topic. However, existing methods usually have three major shortcomings: first, the spatial dependency based on the static adjacency matrix is difficult to reflect the real-time evolution of the road network; second, the spatiotemporal features are often modeled separately, and deep coupling cannot be achieved; third, the model structure complexity is high, which does not match the scale of small sample data and easily leads to overfitting. To this end, this paper proposes LCA-STGNet (Link-Coefficient Adaptation Spatio-Temporal GNN), which uses LSTM as a temporal encoder, uses a dynamic adjacency matrix to capture time-varying spatial features, and enhances information interaction and expression capabilities through cross-node attention mechanisms and residual connections. On a multi-section traffic speed dataset, LCA-STGNet performs significantly better than HA, ARIMA, LSTM, GCN+LSTM, and STGCN in multi-step prediction: compared with LSTM, the average RMSE is reduced by $41.5 \%$; compared with STGCN, MAE is reduced by $10.1 \%$ and $\mathbf{R}^{\mathbf{2}}$ is increased by 4.3 percentage points; and it remains robust in long sequence predictions, verifying its robustness and generalization ability in small sample scenarios.