Traffic State Identification Based on Graph Neural Networks: Taking Tidal Traffic as an Example.

Traffic state identification is an important step before deploying many applications in intelligent transportation systems. Under the background of traffic big data and internet of things, traditional methods are unable to utilize heterogeneous data or deal with spatial-temporal correlations. Methods based on Graph Neural Network (GNN) are promising solutions to these issues, but are rarely used. This study applied a GNN-based model to the tidal traffic route identification. The traffic state identification was modeled as a node classification problem. A novel graph representation of road networks based on flow turns was proposed to represent heterogeneous data and tackle data missing issues. The Graph Attention Network (GAT) and Gate Recurrent Unit (GRU) were adopted to model the spatial-temporal correlations. Compromising on the difficulty of diverse real data collection, the model was trained with synthetic data and tested with real data. With tidal traffic as an example, results showed the good adaptabilities of GNN in traffic state identification, and the outperformance of proposed graph structure against two traditional representations. Overall, this study indicated the practical prospects of GNN in traffic state identification.