Waiting time estimation for ride-hailing fleets using graph neural networks

Ride-hailing services are part of intermodal transport systems, allowing passengers to use various transport modes for their trip. The optimal choice for a request in the intermodal system depends on the passenger's waiting time for the ride-hailing service. Estimating this waiting time is crucial for efficient system operation. The prediction of waiting time depends on the spatial dependency of the transport network and traffic flow elements. Graph neural network (GNN) approaches have gained attention for capturing spatial dependencies in various applications, though less attention has been given to ride-hailing waiting time prediction. The aim of this master thesis is to implement a GNN-based method to predict waiting time for ridehailing requests in the network. Simulation-based waiting time data is used for model training and validation. MATSim is chosen for generating waiting time data under different demand and supply scenarios. Graph Convolutional Network (GCN) and Gated Attention Network (GAT) are used as prediction models. Regression and MLP methods are used as baselines to compare model performance. Results show GCN outperforms regression by 15%, while GAT performs 14% better than regression.