Abstract:Traffic forecasting is essential to intelligent transportation systems, which is challenging due to the complicated spatial and temporal dependencies within a road network. Existing works usually learn spatial and temporal dependencies separately, ignoring the dependencies crossing spatial and temporal dimensions. In this paper, we propose DSTCGCN, a dynamic spatial-temporal cross graph convolution network to learn dynamic spatial and temporal dependencies jointly via graphs for traffic forecasting. Specifically, we introduce a fast Fourier transform (FFT) based attentive selector to choose relevant time steps for each time step based on time-varying traffic data. Given the selected time steps, we introduce a dynamic cross graph construction module, consisting of the spatial graph construction, temporal connection graph construction, and fusion modules, to learn dynamic spatial-temporal cross dependencies without pre-defined priors. Extensive experiments on six real-world datasets demonstrate that DSTCGCN achieves the state-of-the-art performance.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to effectively capture the complex spatio - temporal dependencies in the road network in traffic prediction. Specifically, existing works usually learn spatial and temporal dependencies separately, ignoring the dependencies across spatial and temporal dimensions. This paper proposes a Dynamic Spatio - Temporal Cross - Graph Convolutional Network (DSTCGCN), aiming to jointly learn the dynamic spatial and temporal dependencies through the graph structure to improve the accuracy of traffic prediction. DSTCGCN introduces an attention selector based on the Fast Fourier Transform (FFT) to select relevant time steps at each time step and learns the dynamic spatio - temporal cross - dependencies without predefined priors through the dynamic cross - graph construction module. Experimental results show that DSTCGCN achieves state - of - the - art performance on six real - world datasets. ### Main Contributions 1. **FFT - based Attention Selector**: Select relevant time steps at each time step. Based on actual traffic data, it can model the dynamic characteristics of temporal dependencies. Moreover, it can limit the number of temporal neighbors at each time step, reducing computational complexity. 2. **Dynamic Cross - Graph Construction Module**: Fuse time - varying spatial graphs and time - connection graphs in a directed and sparse manner, and can model dynamic spatio - temporal cross - dependencies without introducing the over - fitting problem. 3. **Extensive Experimental Evaluation**: DSTCGCN is evaluated on six real - world datasets, and the results prove its state - of - the - art performance in traffic flow and speed prediction. ### Technical Details - **FFT - based Attention Selector**: - Utilize Time Normalization (TN) to extract high - frequency components of traffic data, reflecting the changing characteristics of the input. - Concatenate these features with the original traffic data and input them into the FFT - based attention selector. - Convert the input into Query (Q) and Key (K) through linear projection, and calculate the attention weights using the Hadamard product in the Fourier space. - Further reduce the computational complexity through mean aggregation, and select the top τ relevant time steps at each time step. - **Dynamic Cross - Graph Construction Module**: - **Spatial Graph Construction**: Combine node embeddings and time embeddings to generate a dynamic spatial graph. - **Time - Connection Graph Construction**: Based on the self - dependencies of the spatial graph and the relevant weights of the selected time steps calculated by the FFT - based attention selector, construct a time - connection graph. - **Fusion**: Fuse the spatial graph and multiple time - connection graphs into a cross - graph, considering the information propagation direction from the past to the future. - **Graph Convolution**: - Introduce cross - graph convolution to extract dynamic spatial and temporal features. - Generate convolution parameters through the idea of decomposition to reduce model parameters. - **Prediction Module**: - Use GRU as the backbone model, and replace the MLP layer in GRU with a graph convolution layer. - Predict traffic data for the next H time steps through linear transformation of the final hidden state. ### Experimental Setup - **Dataset**: Evaluate on six publicly available traffic datasets, including PEMS03/4/7/8 and METR - LA/PEMS - BAY. - **Experimental Setup**: The data is divided into training, validation, and test sets in chronological order. The input is one - hour data, and the output is the next - hour data. - **Evaluation Metrics**: Include Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE). ### Baseline Methods - **Classical Time - Series Prediction Methods**: Such as HA, ARIMA, VAR, and FC - LSTM. - **STGNN with Static Graph**: Such as STGCN, ASTGCN, Graph WaveNet, AGCRN, MTGNN, StemGNN, Z - GCNETs, and RGSL. Through these techniques, DSTCGCN effectively solves the problem of complex spatio - temporal dependencies in traffic prediction, improving the accuracy and robustness of prediction.

DSTCGCN: Learning Dynamic Spatial-Temporal Cross Dependencies for Traffic Forecasting

DMGSTCN: Dynamic Multi-Graph Spatio-Temporal Convolution Network for Traffic Forecasting

Spatial–Temporal Dynamic Graph Convolutional Network With Interactive Learning for Traffic Forecasting

Spatial–Temporal-Correlation-Constrained Dynamic Graph Convolutional Network for Traffic Flow Forecasting

Spatio-Temporal Graph Convolutional Networks: A Deep Learning Framework for Traffic Forecasting

Graph learning-based spatial-temporal graph convolutional neural networks for traffic forecasting

CDGNet: A Cross-Time Dynamic Graph-based Deep Learning Model for Traffic Forecasting

Dynamic Frequency Domain Graph Convolutional Network for Traffic Forecasting

Dual Dynamic Spatial-Temporal Graph Convolution Network for Traffic Prediction

Spatial-Temporal Interactive Dynamic Graph Convolution Network for Traffic Forecasting

3D Graph Convolutional Networks with Temporal Graphs: A Spatial Information Free Framework For Traffic Forecasting

Decoupled Dynamic Spatial-Temporal Graph Neural Network for Traffic Forecasting

ISTD-GCN: Iterative Spatial-Temporal Diffusion Graph Convolutional Network for Traffic Speed Forecasting

Attention-based spatial–temporal adaptive dual-graph convolutional network for traffic flow forecasting

Spatial dynamic graph convolutional network for traffic flow forecasting

Spatial–Temporal Complex Graph Convolution Network for Traffic Flow Prediction

Spatio-Temporal Joint Graph Convolutional Networks for Traffic Forecasting

Predicting Traffic Flow Using Dynamic Spatial-Temporal Graph Convolution Networks

KST-GCN: A Knowledge-Driven Spatial-Temporal Graph Convolutional Network for Traffic Forecasting

Adaptive Spatio-Temporal Convolutional Network for Traffic Prediction.

Attention Based Spatial-Temporal Graph Convolutional Networks for Traffic Flow Forecasting