Ride-sharing is a critical challenge within intelligent transportation systems, playing a vital role in various applications such as demand-responsive transit. Traditional methods, including heuristic algorithms and mathematical models, often require substantial computation time, limiting their practicality in dynamic environments. In recent years, machine learning approaches have emerged as promising alternatives, offering more efficient solutions through offline training followed by online deployment. However, many existing studies have neglect the time-varying nature of transportation networks, which can considerably affect route efficiency and travel times. To address this issue, this paper proposes an innovative encoder-decoder model that incorporates asymmetric attention weights. During the encoding phase, the model captures the distance relationships between nodes, effectively representing the structure of the road network. During the decoding phase, it realizes rapid generation of vehicle routes considering the road network structure. Comprehensive experiments conducted on real-world road networks validate the model’s effectiveness. The results demonstrate that it not only produces high-quality routing solutions but also operates with greater efficiency compared to traditional methods.