TransLGX: A Self-Contained Model to Predict the Entire Lifecycle and Complete State of Logistics Package Trajectories

This paper presents a novel logistics package trajectory prediction approach. This approach breaks away from the traditional Markov assumption for the first time by capturing the dynamic changes of the logistics network from previous trajectory events and covers the entire lifecycle and complete state of logistics package trajectories. Furthermore, based on this approach, we designed an industrial-grade prediction model, TransLGX. TransLGX captures previous trajectory within the latent space via causal attention, achieving self-containment and high robustness. It then employs an autoregressive architecture and a multi-scale address encoder to achieve highly accurate predictions of the entire lifecycle of packages from creation to delivery. Finally, TransLGX utilizes a two-stage multi-output head to jointly predict the complete state of the trajectory, including actions, locations, transit durations, and delivery time. The self-containment, high robustness, and high accuracy of TransLGX enhance the model's industrial-level practical applicability. Our extensive experiments on datasets comprising tens of millions of package trajectories from six countries demonstrate the effectiveness of TransLGX. We compare against multiple state-of-the-art baseline models from recent years, ensuring a rigorous and comprehensive evaluation. The billion-parameter TransLGX consistently achieves superior performance across all four complete state prediction tasks. Furthermore, TransLGX has been successfully deployed on the TikTok shop e-commerce platform, assisting millions of users in tracking their packages. This successful application not only validates the practical value of our model but also highlights the promise of non-Markovian sequence modeling methods, offering new research directions in this domain. Supplementary materials are available at: https://github.com/Scholar-Song/TransLGX/blob/main/