【作者】 赵杰；

【导师】 陈超；

【作者基本信息】 重庆大学 ， 计算机科学与技术， 2024， 博士

【摘要】 随着城市现代化的加速及移动出行服务的迅速发展,城市交通系统正承受着前所未有的压力。不断增长的人口规模使得城市交通拥堵等问题日益突出,新型移动出行服务(如滴滴和Uber)的兴起深刻地改变了人类的出行模式,加剧了交通系统的复杂性。在此背景下,路网作为城市交通系统的核心组成部分,扮演着至关重要的角色。路网既是支撑人类移动出行的基础设施,也是蕴含丰富语义的地理空间实体,这种双重特性赋予了路网广泛的研究价值。一方面,针对路网时空数据的预测学习可以帮助理解城市交通动态,为交通管理和出行规划提供合理参考。另一方面,针对路网及其丰富语义的表示学习可以构建高质量的路网向量表示,为智能交通系统的多种应用提供支持。近年来,移动感知技术的发展和人工智能技术的进步,为路网的预测及表示学习研究提供了丰富的数据资源和坚实的理论基础。然而,在现实的路网场景中,这一研究仍然面临几大关键挑战:1)路网中的时空数据(如交通流量)存在着多维度、动态的时空依赖,难以准确建模;2)路网中的多模式需求(如出租车和网约车需求)存在着复杂的时空关联,难以精准解耦;3)路网中蕴含着复杂的地理空间语义和移动性语义,难以有效表征。针对上述挑战,本文从灵活多变时空依赖下的交通流预测,多模式交通需求耦合下的联合预测,复杂语义下的路网表示学习等三个方面展开研究。具体来说:首先,针对灵活多变的时空依赖难建模挑战,本文以交通流预测任务为导向,提出了一种双图门控循环卷积神经网络模型(Dual Graph Gated Recurrent Neural Network,DG~2RNN)。在该模型中,设计了一个双图卷积模块从道路距离和时变相关性两个角度捕捉交通流的局部空间依赖;提出了一个双向的门控循环层学习交通流演化的上下文信息,并采用时间注意力机制评估不同历史时刻流量变化的重要性;设计了一个空间注意机制学习任意两个位置之间的隐含相关性以捕捉交通流的全局空间依赖。在三类交通数据集上的实验结果表明了DG~2RNN模型在交通预测任务上的有效性。其次,针对多模式需求的时空关联难解耦挑战,本文以多模式需求预测任务为导向,提出了一种时空交织神经网络模型(Temporal Spatial Intertwined Network,TSIN)。在该模型中,设计了两个独立的出租车模式组件和网约车模式组件,通过多个堆叠的时空卷积块捕捉模式内的时空依赖;提出了一个时间交织模块和一个空间交织模块,为两侧的模式组件提供信息共享的通道以捕捉模式间的时空依赖。在四个出行需求数据集上的实验结果表明了TSIN模型在需求联合预测任务上的有效性,并通过进一步融入共享单车的需求预测任务验证了模型的可扩展性。最后,针对语义复杂的道路网络难表征挑战,本文以路网表示学习任务为导向,提出了一种语义增强的图对比学习框架(Semantic-Enhanced Graph Contrastive Learning,SE-GCL)。在该框架中,设计了一个多模态特征嵌入模块,利用道路属性和街景图像丰富路段的初始特征;提出了一个语义增强的图增广模块,通过基于移动性的边删除和基于模态的特征遮盖操作生成两个不同的路网视图;设计了一个语义增强的对比优化模块,利用路网的地理空间语义和移动性语义选择正负样本,并从视图内和视图间两个方面进行对比优化。在两个城市路网的三个下游应用的实验结果表明了SE-GCL框架在学习路网表示上的有效性。更多还原

【Abstract】 With the acceleration of urban modernization and the rapid development of mobility services,urban transportation systems are facing unprecedented pressures.The growing population scale leads to a series of traffic problems,while the emergence of new mobility services(e.g.,Didi and Uber)profoundly changes human mobility patterns,further exacerbate the complexity of the transportation system.In this context,the road network,as a core component of urban transportation systems,plays a crucial role.Specifically,road networks serve as both the infrastructure supporting human mobility and a geographic entity rich in semantics.This dual nature endows road networks with extensive research value.On the one hand,predictive learning on spatiotemporal data of the road network can help understand urban traffic dynamics and provide reasonable references for traffic management and travel planning.On the other hand,representation learning of the road network with rich semantics can construct high-quality representations,supporting various applications of intelligent transportation systems.In recent years,the development of mobile sensing technology and the advancement of artificial intelligence have provided abundant data resources and solid theoretical foun-dations for predictive and representation learning on road networks.However,several key challenges remain in real-world road network scenarios:1)The spatiotemporal data(i.e.,traffic flow)in the road network presents multi-dimensional and dynamic spatiotemporal dependencies,which are difficult to model accurately;2)Multi-mode demands in the road network(such as taxi and ride-hailing demands)exhibit complex spatiotemporal corre-lations,which are difficult to decouple precisely;3)The road network entails complex geographical and mobility semantics,which are difficult to characterize effectively.To address these challenges,this dissertation investigates predictive learning on traffic flow under flexible and dynamic spatiotemporal dependencies,joint prediction of multi-mode demands under coupled spatiotemporal dependencies,and representation learning for road networks under complex semantics.Specifically:Firstly,to tackle the challenge of modeling flexible and varying spatiotemporal dependencies,this dissertation proposes a Dual Graph Gated Recurrent Neural Network(DG~2RNN)model oriented to the task of traffic flow prediction.In the model,a dual graph convolution module is designed to capture the local spatial dependencies of traffic flow from the perspectives of road distance and time-varying correlations.A bidirectional gated recurrent layer is introduced to learn the contextual information of traffic flow evolution,and a time attention mechanism is adopted to evaluate the importance of traffic flow changes at different historical moments.Additionally,a spatial attention mechanism is designed to learn the implicit correlations between any two locations for global spatial dependencies modelling.Experimental results on three types of traffic datasets demonstrate the effectiveness of the DG~2RNN model.Secondly,to address the challenge of decoupling the spatiotemporal correlations of multi-mode demands,this dissertation proposes a Temporal Spatial Intertwined Network(TSIN)model oriented to the task of multi-mode demands co-prediction.In the model,two independent mode components for taxi and ridesourcing services are designed,with multiple stacked temporal-spatial convolution blocks capturing the spatiotemporal de-pendencies within each mode.A temporal intertwined module and a spatial intertwined module are presented to provide information-sharing channels between the two mode components to capture the spatiotemporal dependencies between modes.Experimental results on four demand datasets demonstrate the effectiveness of the TSIN model in multi-mode demands co-prediction.Further integration of bike-sharing demand prediction task validates the scalability of the model.Finally,to tackle the challenge of characterizing the road network with complex semantics,this dissertation proposes a Semantic-Enhanced Graph Contrastive Learning(SE-GCL)framework oriented to the task of road network representation learning.In the framework,a multi-modal feature embedding module is designed to enrich the initial features of road segments by using road attributes and street-view images.A semantic-enhanced graph augmentation module is proposed to generate two different road network views through edge removing based on mobility and feature masking based on modality.Additionally,a semantic-enhanced contrastive optimization module is presented to select positive and negative samples based on the geographical and mobility semantics of the road network and perform contrastive learning from both intra-view and inter-view perspectives.Experimental results on three downstream applications of two city road networks demonstrate the effectiveness of the SE-GCL framework.更多还原