地理知识图谱增强与多时空条件约束的轨迹预测  

Trajectory prediction enhanced by geographic knowledge graph and multi-spatio temporal constraints

作　　者：李佳[1] 李静[1] 刘海砚[1] 陆川伟 陈晓慧[1] 刘俊楠[2] 石文 LI Jia;LI Jing;LIU Haiyan;LU Chuanwei;CHEN Xiaohui;LIU Junnan;SHI Wen(School of Data and Target Engineering,University of Information Engineering,Zhengzhou 450001,China;School of Geo-Science&Technology,Zhengzhou University,Zhengzhou 450001,China;Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]信息工程大学数据与目标工程学院,河南郑州450001 [2]郑州大学地球科学与技术学院,河南郑州450001 [3]中国科学院空天信息创新研究院,北京100190

出　　处：《测绘学报》2024年第10期2021-2033,共13页Acta Geodaetica et Cartographica Sinica

基　　金：国家自然科学基金(42371438,42301526,62303439);河南省自然科学基金(242300420623)。

摘　　要：基于机器学习的轨迹预测方法通常依赖历史轨迹数据的数量和质量,而社交媒体签到数据更新频率低,形成的轨迹稀疏,在预测中易出现难学习、过拟合等问题。为突破低质量轨迹数据在预测任务中的限制,本文提出一种基于地理知识图谱增强与多时空约束条件建模的轨迹预测方法。将复杂异构的多源地理信息结构化为由若干三元组构成的地理知识图谱进行统一表达,并通过知识表示模型挖掘其中语义关联来增强轨迹序列的向量表征,同时采用具有多重时空约束条件的多头自注意力机制提取稀疏轨迹序列中的多重时空特征,从而提升轨迹预测精度。研究采用纽约市Foursquare社交媒体签到数据进行方法验证,试验结果表明:本文方法相较于其他表示学习方法和轨迹预测方法,在命中率和平均倒数排名两个评价指标上均有不同程度的提升,能够有效增强稀疏轨迹序列的表征,提取轨迹的多重时空特征,提高社交媒体用户签到轨迹的预测精度。Trajectory prediction methods based on machine learning typically rely on the quantity and quality of historical trajectories.But social media check-in data has a low update frequency,that would lead to difficulties in learning and overfitting during trajectory prediction.To overcome the difficulty of low-quality trajectories data in prediction tasks,we propose a trajectory prediction method enhanced by geographic knowledge graph and multi-spatio temporal constraints.The proposed model transforms complex and heterogeneous multi-source geographic information into a geographic knowledge graph composed of several triples for unified expression,and mines entity associations through knowledge embedding models to enhance the feature representation of trajectories.At the same time,the model utilizes a multi-head self-attention with multi-spatio temporal constraints to extract multiple features from check-in trajectories.The proposed model is validated on Foursquare social media check-in data from New York.It s shown in experiment results that proposed model has improved to varying degrees in hit rate(HR)and mean reciprocal rank(MRR)evaluation indicators,comparing with other representation learning methods and prediction models.The result indicate that the proposed model can effectively enhance the representation of check-in trajectories,extract multiple temporal features of trajectories,and improve the prediction accuracy of social media user check-in trajectories.

关 键 词：轨迹预测 多源地理信息 社交媒体 地理知识图谱 多时空约束 

分 类 号：P208[天文地球—地图制图学与地理信息工程]