Taxi Destination Prediction Achieves Improved Accuracy with Hybrid Quantum-Classical Networks

Predicting the destination of urban taxis presents a significant challenge for modern city planning and traffic management, and researchers are now exploring the potential of quantum computing to improve accuracy. Xiuying Zhang, Qinsheng Zhu, and Xiaodong Xing propose a novel approach, a Spatio-Temporal Hybrid Quantum-Classical Graph Convolutional Neural Network, which combines the strengths of both classical and quantum computation. This new method leverages graph convolutional networks to encode road network features, then maps these onto quantum circuits, while simultaneously modelling temporal dependencies using classical theory. The team demonstrates that this hybrid architecture outperforms existing prediction methods, suggesting that integrating quantum computation offers a powerful new tool for understanding and optimising urban mobility.

Scientists aim to leverage the potential of quantum computing to improve traditional graph convolutional networks used for this task. The proposed model combines classical GCN layers with quantum layers, designed to process graph information using quantum principles and potentially capture more complex relationships than classical methods. The team successfully integrated classical and quantum computing techniques, combining graph convolutional networks and temporal convolutional networks with a quantum graph convolutional module and a building-oriented contextual semantic vector. Through this hybrid approach, the algorithm effectively captures both spatial dependencies within the road network and the dynamic evolution of trip patterns, leading to demonstrably improved accuracy and stability in predicting taxi destinations. Experimental results across three cities, Porto, San Francisco, and Manhattan, consistently demonstrate the superiority of H-STQGCN compared to existing methods and purely classical networks.

The algorithm achieves significant reductions in prediction error, with notable improvements observed in the complex urban landscape of Manhattan, where a 14. 1% relative improvement in root mean squared error was recorded. The study pioneers a dual-branch architecture designed to capture intricate patterns in trajectory data. This classical component is then augmented by a quantum module, which maps these graph features onto parameterized quantum circuits, enabling the exploration of exponentially larger feature spaces. The time evolution branch addresses the dynamic aspects of taxi trips by integrating multi-source contextual information and capturing dependencies using a classical Temporal Convolutional Network. The system fuses the outputs of the spatial and temporal branches, creating a holistic representation of each trip’s context and modeling both the static road network topology and the dynamic flow of traffic. The team proposes a hybrid spatial representation framework that combines classical road network topology extraction with quantum high-dimensional feature mapping, enhancing the ability to model complex road network structures. These layers map graph features onto parameterized quantum circuits, enabling deeper feature modeling of graph-structured data via quantum state evolution. The research introduces a collaborative fusion mechanism that integrates extracted quantum spatial features with temporal dynamic modeling, achieving unified characterization of non-local spatial dependencies and temporal evolution patterns. This approach leverages the high-dimensional mapping advantages of quantum computing to compensate for shortcomings in classical deep learning, constructing an end-to-end solution for vehicle destination prediction.