【作者】 陈宇；

【导师】 秦昆；

【作者基本信息】 武汉大学 ， 地图学与地理信息系统， 2023， 硕士

【摘要】 场所(Place)承载了人类的行为活动,包含了空间特征和人类行为活动相关的特征,是结合人群视角理解地理环境的重要途径之一,可以被具象为兴趣点、地理格网、城市、区域等地理研究单元。场所表征学习可以将场所表示为数字化的、包含了场所属性特征的低维稠密特征向量,是许多数据驱动的城市研究和应用的初步任务,可以用于揭示场所内在规律,提升位置预测等下游任务性能,对驱动地理人工智能(Artificial intelligence in geography,Geo AI)的发展有重要意义。场所的属性包括两个方面:(1)场所内部和自身的一阶属性特点(一阶量);(2)场所间动态的空间邻近交互和行为邻近交互作用产生的二阶属性特点(二阶量)。不同模态的数据从不同视角表示场所的属性特点,比如兴趣点数据和图像数据可以用于建模场所功能、内部地物形态信息等一阶量,轨迹数据可以用于建模场所间动态交互的二阶量,如空间自相关性、空间互补性等。以往研究通常仅关注轨迹中停留点之间的交互关系(比如出发地与目的地),忽略了实现这种停留点间行为邻近交互的细粒度轨迹,忽视了其中刻画的多级空间距离约束及空间邻近交互作用;此外,直接将特定领域的表征建模方法应用到地理学领域难以对场所的地理特性(如时空特性)进行较好的建模;第三,目前研究大多采用联合训练的结构融合不同模态的信息,导致模型依赖所有模态的信息作为输入,所得的场所表征无法应用于只有单模态或缺失部分模态的场景,模型的可复用性有限。本论文设计了一个多模态对比式预训练场所表征模型(Multi-Modal Contrastive Pre-training Model for Place Representation,MMCPP),在时空框架下,以数据驱动的方式,将兴趣点数据(Point-of-Interest,POI)、图像数据和轨迹数据所表示的场所的一阶属性和二阶属性建模到场所表征中。将地理格网作为场所具象后的地理研究单元,以武汉市三环线矩形包络为研究区范围,采集相关数据实现了本文所提出的模型,并开展了定性和定量实验,验证了模型的有效性。本文的主要研究工作和成果如下:(1)建模各模态表示的场所属性本文采用了3个编码器及相应的3类地理大数据,并使用相应的自监督预训练任务,将各个模态表示的场所属性特征建模到场所表征中。其中,POI编码器和图像自编码器建模了场所的一阶属性,POI编码器用于从POI的类别分布和地理位置分布中提取场所的功能属性;图像自编码器用于从图像中捕获地物的大小、形状、空间位置分布等特征。轨迹编码器Rodt Former以细粒度的时空轨迹序列作为输入,其结合自注意力机制以及本文所设计的旋转时间差位置编码,建模了空间邻近交互和行为邻近交互作用以及动态交互的时间特性,以表示场所之间的二阶属性(如空间自相关性、空间互补性等)。(2)多模态协同预训练融合各模态表示的场所属性多模态的地理大数据从多个视角表示场所的语义,因此融合不同模态信息的表征可以更加全面地刻画场所属性。本文引入了一个多模态协同预训练阶段,在利用兴趣点数据和图像数据建模场所内部及自身的一阶属性、轨迹数据刻画场所之间动态交互作用所产生的二阶属性的基础上,提出了一种面向场所表征建模的对比样本构建过程,结合对比学习方法,进一步通过协同训练的方式融合不同模态表示的场所属性,最终的场所表征模型包含三个可独立编码的编码器组件,可以在融合POI、图像和轨迹数据中表示的场所语义的同时,适应单模态或缺失部分模态的场景,提高了模型的可复用性。(3)多模态对比式预训练场所表征模型有效性验证本文将地理格网作为场所具象后的地理研究单元,以武汉市三环线矩形包络为研究区范围,收集了相关的POI数据、地图切片数据以及出租车轨迹数据作为各个模态建模的数据源,结合定性和定量的方法,开展实验验证了模型的有效性。首先,采用对3个编码器组件进行了定性分析,验证了单模态预训练组件的表征能力和表征可靠性。之后,引入“下n个时间切片位置预测”这一下游任务,设计了对比实验,通过比较本文方法和其他相关模型在下游任务上的指标,定量验证了本文所提出模型的泛化性和表征能力。最后,通过设计消融实验,验证了模型的轨迹编码器组件中所设计的旋转时间差位置编码的有效性,以及模型融合不同模态的有效性,并讨论了模型在训练过程中采样轨迹数量这一超参数对模型建模场所表征效果的影响。实验结果表明:相比基准方法,本文模型中所设计的旋转时间差位置编码以及融合多模态的信息,都使得模型在预训练过程中建模了更高质量的场所表征,在一定程度上可以帮助下游任务模型实现更好的性能。更多还原

【Abstract】 The place contains attributes related to spatial features and human behavior and is one of the important ways to understand the geographical environment from the perspective of human beings.Place representation learning can represent the place as a numerical low-dimensional dense vector,which is an initial task for many data-driven urban studies and applications.And the dense vector can be used to reveal the internal laws of the place and improve the performance of downstream tasks,which is of great significance to promote the development of artificial intelligence in geography(Geo AI).The attributes of a place include two aspects:(1)the first-order attributes of the place itself;(2)the second-order attributes are characterized by spatial proximity interaction and behavioral proximity interaction between places.Data of different modalities represent the attributes of the place from different perspectives.For example,point-of-interests(POI)and images can be used to represent first-order attributes such as place functions and surface information,and trajectories can be used to simulate interactions between places to describe second-order attributes such as spatial autocorrelations and spatial complementarity.There are some deficiencies in the existing related research.First,previous studies usually focus on the interaction between stay points of trajectories(such as pick-up and drop-off points),while ignoring the fine-grained trajectory sequences which contain multi-level spatial distance constraints and spatial proximity interactions.Second,representation modeling methods in the field of computer science are difficult to express the geographical features of the place,such as spatial and temporal features.Third,most current studies employ the joint representation structure to integrate the information of different modalities,resulting in models that rely on all modalities as input.It means that the model cannot be applied to tasks where there is only one modality or part of the modalities are missing,and the reusability of the model is limited.To fill the research gap,this paper proposes a Multi-Modal Contrastive Pretraining model for Place representation(MMCPP)and conductes qualitative as well as quantitative experiments to verify the effectiveness of the model.The main research contents of this paper are as follows:(1)Encoding of place attributes represented by each modalityMMCPP contains three encoders that correspond to three modalities of data to describe the place attributes by using the self-supervised pre-training tasks respectively.A POI encoder extracts the functional attributes of the place from the category distribution and geographical location distribution from the POI data.And an image autoencoder captures surface information such as size,shape,and spatial position distribution of ground objects from images.Both of them encode the first-order attributes of the place.Additionally,this paper proposed a trajectory encoder,named Rodt Former,which takes fine-grained spatiotemporal trajectories as input,and used the self-attention mechanism with the rotary temporal interval position embedding,to mine the second-order attributes of the place(such as spatial autocorrelations and spatial complementarity).These second-order attributes are influenced by the dynamic interactions of spatial proximity and behavioral proximity between places.(2)Coordinated integration of place attributes from each modal representationMulti-modal data represents the semantics of places from multiple perspectives,and the integration of different modal information can more comprehensively describe the attributes of the place.This paper designed a multi-modal coordinated pre-training stage including a contrastive sample sampling method for place representation modeling based on the contrastive learning method.And then,the place attributes represented by different modalities are integrated by the coordinated representation structure,so that MMCPP can not only integrate the place attributes represented by POI,image,and trajectories but also adapt to the tasks of a single modality or missing part modality.(3)Verification of the effectiveness of MMCPPIn this paper,the rectangular envelope of the Third Ring Road in Wuhan,China,is taken as the research area,and the geographic grid is used as the geographic research unit.Relevant data are collected for each modality in MMCPP,including POIs,maptiles,and taxi trajectories.The model is realized and its effectiveness is verified by qualitative and quantitative experiments.First,the semantics of the place representation encoded by each unimodal pre-trained encoder is analyzed to verify the representation ability and reliability of each encoder.Second,the downstream task of "location prediction at next n times" is introduced to quantitatively evaluate the generalization ability of MMCPP.Finally,the ablation experiments are designed to verify the effectiveness of the rotary temporal interval position embedding in the trajectory encoder and the model integration of different modalities.Additionally,the impact of the number of sampled trajectories during model training is discussed.The experimental results show that: the design of the rotary temporal interval position embedding and the integration of multi-modal information included in MMCPP enables the model to build a higher quality place representation during the pre-training process,which can help downstream task models achieve better performance compared with baseline methods in a certain extent.更多还原