Analysis of Different Graph Convolutional Network Prediction Models with Spatial Dependence Evaluation

Over the past few years, the graph-based Graph Convolutional Network (GCN) models have been introduced to handle the spatial-temporal attributes in short-term traffic prediction. Aiming at filling the gap of lacking comparisons of different graph-based models, this study analyses the performances of different graph-based convolutional neural networks. A modified sequence to sequence structure with attention mechanism and the residual module is applied as the backbone of the prediction model, and the input to the graph-based models is the Maximal Information Coefficient (MIC) adjacency matrix. The results confirm the effectiveness of the proposed prediction framework and shed light on choosing the most acceptable graph-based convolutional neural networks for short-term traffic predictions.