广义时空数据分析与挖掘

Recent years have witnessed the rise of data-driven management and decisions. To handle the variety characteristic of big data, this project first provides a unifying representation framework for data objects belonging to different types. Based on the unifying representation, this project explores the relationships of different data types, such as time series, temporal sequences, images, texts, classical spatiotemporal data, etc., and also points out the importance of Generalized SpatioTemporal Data (GSTD). In this project, we focus on two types of GSTD: lattice and graph/network GSTD. For lattice GSTD, this project aims to provide a novel deep spatiotemporal neural network to realize the decomposition and fusion of lattice GSTD with consideration of spatial and temporal association. For graph/network GSTD, this project targets at the development of a novel dynamic graph convolutional neural network to capture the structure of graph/network and to model the dynamics of graph/network evolution. This project will also study the exclusive multi-task learning mechanism to process the generalized spatiotemporal relationships of multiple outputs. The proposed methods will be used in many important smart-city or sharing-economy problems. The deep spatiotemporal neural network will be used for co-prediction of multiple traffic demands, which is able to present the next-period prediction of “subway/bus-taxi-bikes” simultaneously, and also improve the overall prediction accuracy by leveraging interactions of multiple traffic modules. The graph convolutional neural network will be used to model the traffic flows of urban road networks, and analyze the evolution of human mobility. The multi-task spatiotemporal learning will be used to further improve the performance of the above prediction systems by transferring knowledge among outputs in a spatiotemporal manner. In summary, this project explores one fundamental issue in data science theory, studies two new types of spatiotemporal data, provides novel data mining tools for them, and applies the proposed methods in several important management and operation problems in smart-city or sharing-economy. Therefore, this project has important theoretical and practical value.

近年来，数据驱动的管理与决策已经渐渐成为学术界和工业界的研究热点。本项目围绕数据科学的基础理论，针对大数据4V特征中的数据类型多样性，率先提出了数据对象的统一化的表示与描述方法，分析了不同类型的数据对象之间的关系，并指出广义时空数据的重要性。围绕晶格型广义时空数据和图/网络型广义时空数据，本项目拟开发深度时空学习网络，实现晶格型广义时空数据的分解与融合；开发动态图卷积神经网络，用于图/网络型广义时空数据的建模与预测；研究面向广义时空约束的多任务学习机制，用于输出端广义时空数据的分析与处理。本项目拟将上述研究成果应用于城市多模态交通需求量的联合预测，城市居民流动性建模分析等智慧城市和共享经济领域的核心商业问题，为城市公共交通规划、共享单车运作管理、出租车路径优化等问题提供数据依据与决策支持。综上，本项目探索了数据科学的基础理论，丰富了数据挖掘的方法工具，具有重要的理论意义和实际应用价值。

针对大数据的多样性问题，提出了广义时空大数据分析与挖掘框架，通过研究数据对象特征之间的相关性，实现了现有绝大多数数据对象的统一化表示与描述。基于该框架，聚焦三类数据对象展开深入研究，提出了相应的模型与方法，包括：1）研究了时态集合这一在真实世界中普遍存在的复杂时序数据，并提出了子图集合表示，基于Transformer的“集合-元素”双层表示学习等方法，在电子商务、AI医疗等领域进行了验证；2）研究了非欧式时空数据的建模与预测方法，提出了空间依赖关系的自适应学习方法，研究了跨模态的空间依赖关系学习，多尺度动态空间邻近关系学习，并将其应用于交通系统运行状态评估。3）研究了通用图表示学习方法，尤其是提出了具有持续时间标签的动态图表示学习方法，以OD交通需求量预测为例展开深入研究，为时空知识图谱的研究提供了新的研究思路。以上研究成果得到了国内外学术界的广泛认可，总共发表学术论文31篇，包括数据挖掘领域国际顶级学术期刊会议TKDE和KDD，以及AAAI、SIGIR、WWW、IJCAI、PR等人工智能和机器学习国内外著名会议和期刊。基于该项目研究成果与交通运输部路网监测与应急处置中心等国家和省市级业务部门联合攻关，提出了时空大数据驱动的交通路网运行状态评估与预测技术，获得2021年度中国公路学会科学技术奖特等奖。培养博士生8名，硕士生10名，多名学生获得国家奖学金，优秀毕业设计论文。