【作者】 张蕾；

【导师】 崔立真；

【作者基本信息】 山东大学 ， 人工智能（专业学位）， 2024， 博士

【摘要】 智慧物流在数字经济中发挥战略性作用。我国在货运量、快递业务量等方面居世界前列,已成为全球最大的物流市场。为深化行业发展,《“十四五”现代物流发展规划》明确指出应强化物流数字化科技赋能,加速数字化转型与智慧化改造。时空计算是智慧物流数字化智能化发展的关键支撑技术之一,其核心是以时空数据为基础,通过深入分析、表示和预测复杂时空模式和趋势,实现精准可靠的智慧物流时空预测,提高物流运输效率和提升客户满意度。随着智能终端的普及和移动大数据的爆炸性增长,智慧物流中的时空计算表现出高度的复杂性,即时空数据关联广泛高阶、时空数据演变频繁多样、时空预测精细化要求高和时空预测不平衡性突出。针对复杂时空计算暴露的高阶时空关系建模难、动态时空表示难、精细物流时空预测难和鲁棒物流时空预测难等一系列问题,结合阿里巴巴供应链智慧物流的实际业务需求,采用时空图和异构超图等计算工具,开展面向智慧物流的复杂时空计算研究,主要研究内容包括:(1)针对复杂物流时空数据间的高阶时空关系建模难的问题,以异构时空超图统一建模,提出了一种面向复杂关联物流数据的时空高阶关系建模方法。该方法将海量物流订单及其多维时空属性建模为物流时空超图,挖掘物流订单间间接的多跳关联关系,实现融合复杂高阶关系的物流时空表示学习,为实现精细可靠的智慧物流时空预测模型提供时空关联建模基础。(2)针对复杂演化物流数据动态时空表示难的问题,以动态图结构学习为核心技术,提出了一种面向演化物流数据的动态时空表示学习方法。该方法将动态物流数据的时空相关性建模为时空图,动态学习并优化生成最优的时空拓扑图结构和动态时空嵌入信息,实现动态感知的物流时空表示学习,为构建精细化的智慧物流时空预测模型提供动态时空表示支撑。(3)针对精细物流时空预测难的问题,以引入多粒度知识的知识蒸馏网络为核心架构,提出了一种面向粗粒度物流数据的精细时空预测方法。该方法在融合高阶关系和动态演化的时空表示基础上,将多粒度的历史物流轨迹构建为多层级物流轨迹图,通过知识蒸馏将轨迹知识蒸馏到粗粒度物流时空嵌入中,捕获细粒度的物流时空信息,实现精细化的物流时空预测,为构建智慧物流鲁棒时空预测模型提供细粒度时空信息支撑。(4)针对鲁棒物流时空预测难的问题,以基于个性化双图的分布识别为核心技术,提出了一种面向不平衡物流数据的鲁棒时空预测方法。该方法在精细物流时空预测的基础上,考虑物流时空数据的不平衡分布特性,设计个性化时空图算法以差异化学习不同频率分布的物流时空数据,获取同时关注高频和低频数据的全局时空分布信息,实现可靠的物流时空预测,构建鲁棒的智慧物流时空预测模型。基于上述创新性研究成果,依托阿里巴巴供应链平台海量真实的物流数据,以智慧物流复杂时空计算方法为支撑,遵循“数据-表示-预测”的研究范式,开发了基于复杂时空计算的智慧物流包裹到达时间预测系统。该系统紧密结合智慧物流包裹到达时间预测的实际应用需求,集成了物流时空数据自动化处理和物流包裹到达时间智能预测两大核心功能,并支持对多种智能预测算法效能的多维度评测,包括预测性能评估、误差分析、鲁棒性分析和系统对比测评。经过实际的多维度测试评估,该系统集成的智能预测算法在预测准确率和物流履约率方面均展现了显著的提升,实现面向复杂物流数据的精细可靠物流时空预测。该系统的集成与实现,不仅有助于理解智慧物流时空数据的复杂关联和动态演变特性,更为智慧物流的精细可靠预测决策提供了坚实的技术支持。更多还原

【Abstract】 Smart logistics plays a strategic role in the digital economy.China has become the world’s largest logistics market,with the volume of freight transport and express delivery services among the world’s largest.To deepen the development of the industry,the "14th Five-Year Plan for Modern Logistics Development" clearly emphasizes the need to strengthen the digital technology empowerment of logistics,and accelerate digital transformation and intelligent upgrading.Spatio-temporal computing is one of the key supporting technologies for the intelligent development of smart logistics.Its core lies in leveraging spatio-temporal data to analyze,represent,and predict complex spatio-temporal patterns and trends,enabling precise and reliable spatio-temporal prediction,improving logistics efficiency,and enhancing customer satisfaction.With the proliferation of smart devices and the explosive growth of mobile big data,spatio-temporal computing exhibits a high level of complexity,i.e.,extensive high-order correlation of spatio-temporal data,frequent and diverse evolution of spatio-temporal data,high requirements for fine-grained spatio-temporal prediction,and prominent imbalance in spatiotemporal prediction.In response to the challenges posed by complex spatio-temporal computing,such as difficulty in modelling high-order spatio-temporal relations,difficulty in dynamic spatiotemporal representation,difficulty in fine-grained spatio-temporal prediction,and difficulty in robust spatio-temporal prediction,we propose to conduct research on complex spatio-temporal computing for smart logistics,taking into account the specific business requirements of Alibaba’s supply chain smart logistics.This research leverages computational tools such as spatio-temporal graphs and heterogeneous hypergraphs.The main research contents include:(1)To address the challenge of modeling high-order spatio-temporal relations in complex logistics data,a method for modeling complex spatio-temporal correlations is proposed.This method utilizes a unified modeling approach based on heterogeneous spatio-temporal hypergraphs.It models massive logistics orders and their multi-dimensional spatio-temporal attributes as hypergraphs,capturing indirect multi-hop correlations among logistics orders,enabling the spatio-temporal representations that integrate complex high-order relations.This provides a foundation for logistics spatio-temporal prediction.(2)To address the challenge of dynamic spatio-temporal representation in logistics data,a method for dynamic spatio-temporal representation learning is proposed.This method is based on the unified modelling of dynamic graph structure learning,which models the spatio-temporal correlation of dynamic logistics data as a spatio-temporal graph,dynamically learns the optimal spatio-temporal topology structure and dynamic spatio-temporal representation,realizes the learning of dynamically-aware spatio-temporal representation,and provides the dynamic representation support for construction of fine-grained spatio-temporal prediction models.(3)To tackle the challenge of fine-grained spatio-temporal prediction in coarse-grained logistics data,a fine-grained spatio-temporal prediction method is proposed.This method combines high-order relations and dynamic evolution in spatio-temporal representation,and incorporates multi-granularity knowledge through knowledge distillation.The method constructs a multi-level trajectory graph using multi-granularity historical logistics data,capturing comprehensive and detailed spatio-temporal representation,enabling accurate and fine-grained spatio-temporal prediction,providing fine-grained representation support for the construction of robust spatio-temporal prediction models.(4)To tackle the challenge of robust spatio-temporal prediction in imbalanced logistics data,a robust spatio-temporal prediction method is proposed.On the basis of fine-grained spatio-temporal prediction,the method considers the imbalanced distribution of data.It designs personalized spatio-temporal graph representation learning that handles different distributed data,capturing differential representations that simultaneously address high-and lowfrequency data,achieving accurate and reliable spatio-temporal prediction,and constructing a comprehensive fine and robust intelligent logistics spatio-temporal prediction model.Based on the above innovative research results,the intelligent logistics package arrival time prediction system based on complex spatio-temporal computing is developed,relying on the massive real logistics data of Alibaba supply chain platform and supported by logistics spatio-temporal computing algorithms.The system is closely integrated with the actual application requirements of package arrival time prediction,integrates two core functions of automated logistics spatio-temporal data processing and intelligent package arrival time prediction,and conducts multi-dimensional evaluation of the effectiveness of intelligent prediction algorithms,including prediction performance evaluation,error analysis,robustness analysis,and system comparison evaluation.Through extensive multi-dimensional testing and evaluation,the integrated spatio-temporal prediction algorithms have demonstrated significant improvements in prediction accuracy and fulfillment rate.The integration and implementation of the system not only contributes to a comprehensive understanding of the complex spatiotemporal data,but also provides solid technical support for precise and reliable prediction decision-making.更多还原