Abstract: Abstract: Population mobility between cities was an essential process of modern urbanization. Identifying the impact of urban development level on population migration played a significant role in optimizing urban industrial structure and socio-economic development. To address the interactive influence between urban characteristics and the strong dependence on historical population mobility data, a High-Order Feature Migration Flow Network（HighFlowNet）model was constructed. Multi-dimensional characteristics of cities were integrated through deep neural networks, and high-order features were extracted by incorporating economic, transportation, and population factors. A dynamic feature enhancement mechanism was employed to highlight key features, while the correlation strength between cities was calculated through global interaction modeling to optimize the prediction of mobility distribution. The model was trained using local urban data and extended to the entire country, enabling a more accurate prediction of inter-city population mobility. The results show that, compared with the extended gravity model (EGM), the HighFlowNet model reduces the root mean square error and mean absolute error by 64.15% and 58.91%, respectively, and improves the consistency of commuter prediction by 13.07%, demonstrating superior predictive performance. The interpretation of model features indicates that under the joint influence of multiple urban characteristics, urban Gross Domestic Product has a positive effect on population mobility, whereas an increase in the proportion of primary industry inhibits inter-city population mobility.

Key words: intercity population flow, Urban multidimensional characterization, higher-order features, urban GDP, percentage of primary sector

摘要： 摘 要: 城市间人口流动是现代城市化的一个重要进程。有效识别城市发展水平对人口迁移影响，对优化城市产业结构和社会经济发展具有重要作用。针对城市特征间交互影响、模型对历史人口流动数据依赖较强的问题，提出一种高阶特征迁移流网络（HighFlowNet）模型。模型通过深度神经网络融合城市的多维特征，结合经济、交通、人口等因素提取高阶特征。采用动态特征增强机制突出关键特征，并通过全局交互建模计算城市间的关联强度，优化流动分布预测。模型基于局部城市数据训练，并推广至全国，实现更精准的城际人口流动预测。结果表明，与扩展重力模型（EGM）相比，HighFlowNet模型在均方根误差和平均绝对误差上分别降低了64.15%和58.91%，在通勤者预测一致性上提升了13.07%，展现较好预测能力。模型特征解释显示，在城市多维特征共同影响下，城市生产总值对城市人口流动有着积极作用，而第一产业占比的提高抑制了城市间人口流动。

关键词: 城际人口流动, 城市多维特征, 高阶特征, 城市GDP, 第一产业占比