DSTIGFN: Dynamic Spatio-Temporal Interactive Graph Fusion Network for Traffic Flow Prediction

Accurate traffic flow prediction is crucial for the efficient management of urban transportation systems. However, existing models generally focus on modeling continuous traffic sequences, ignoring the multilevel correlation characteristics of spatio-temporal heterogeneity in traffic networks at different time scales. Moreover, the spatial dependencies among road nodes are usually modeled by static graphs, making it difficult to capture their dynamic evolution over time. To address the above challenges, this paper proposes a Dynamic Spatio-Temporal Interactive Graph Fusion Network. First, the model periodically encodes the original traffic flow sequence to extract the periodic patterns and global temporal structures embedded in it. Building on this foundation, a trend-preserving parity decomposition is applied to divide the time series into odd and even subsequences, preserving temporal trends while introducing diverse temporal perspectives to support multi-scale temporal modeling. For each subsequence, adaptive and supra-Laplacian graph structures are constructed to capture evolving spatial dependencies and deep temporal correlations. Graph convolution and a subgraph interaction fusion strategy are then employed to integrate complementary information across subsequences. Finally, by restoring the odd and even subsequence features to their original temporal order, the model effectively integrates global and local spatio-temporal dependencies, thereby enhancing its ability to predict complex traffic patterns. The source code is available athttps://github.com/Zachary-Qi/DSTIGFN-master