Traffic forecasting is one of the important means to alleviate congestion in urban transport systems. Although many works have been devoted to improving the performance of traffic forecasting, the rapid development of urbanization has also brought new challenges, such as (1) the spatial semantic attributes of traffic nodes and regions remain unexplored; (2) the graph structure constructed using short-term historical windows is not reliable; (3) it is difficult to collaboratively capture the long-term trends and semantic events in the time series. To address these challenges, we propose the Hierarchical Spatial-Temporal Semantic-Enhancement Network For Traffic Forecasting (HSTSNet). Specifically, we divide the nodes into region levels through clustering and design a two-layer architecture. During spatial features learning phase, we design a Spatial Semantic Block which enhances graph structure by Discrete Wavelet Transform (DWT). Then, we develop a Temporal Semantic Block with two key parts: 1) a Dilated TCN layer for extracting temporal dependencies; 2) a Semantic-aware Frequency Enhanced Channel Attention Mechanism (Semantic-FECAM) leveraging Discrete Cosine Transform for dynamic frequency decomposition and channel attention fusion. This approach synergizes long-term trend and semantic detail modeling, overcoming challenges in fusing global frequency dependencies with context-specific semantic features in time-series prediction. Finally, we integrate node and region features and make predictions. Experiments on two real-world datasets demonstrate the superior performance of our method over several baselines.
