Understanding urban human mobility, particularly trip purposes, is essential for optimizing traffic management, personalized recommendations, and urban planning. However, in real-world scenarios, trip purposes cannot be directly extracted from the trajectory data. To address this issue, we propose a geometry-driven neural topic model for trip purpose inference. We integrate trajectory data with nearby POI data using a geometry-driven technique to enhance the interpretability of the results. Furthermore, our model captures the semantics and relationships of the data in a high-dimensional space and identifies latent topics representing distinct trip purposes. These learned topics are analyzed using clustering algorithms to group similar trips, enabling trip purpose inference. And we evaluate our model using the trajectory data of Shenzhen and Chengdu, and compare it with baseline models. The results demonstrate that our model performs well. Furthermore, we analyze trajectory data containing trip purpose information to gain insight into human mobility patterns and the influence of trip purposes, paving the way for potential implications and future research directions.

了解城市人员流动，特别是出行目的，对于优化交通管理、个性化建议和城市规划至关重要。然而，在现实场景中，无法直接从轨迹数据中提取出行目的。为了解决这个问题，我们提出了一种用于旅行目的推断的几何驱动的神经主题模型。我们使用几何驱动技术将轨迹数据与附近的 POI 数据集成，以增强结果的可解释性。此外，我们的模型捕获高维空间中数据的语义和关系，并识别代表不同旅行目的的潜在主题。使用聚类算法对这些学习的主题进行分析，以对类似的旅行进行分组，从而实现旅行目的推断。我们使用深圳和成都的轨迹数据评估我们的模型，并将其与基线模型进行比较。结果表明我们的模型表现良好。此外，我们分析包含出行目的信息的轨迹数据，以深入了解人类流动模式和出行目的的影响，为潜在影响和未来研究方向铺平道路。