【作者】 刘洋；

【导师】 刘志远；

【作者基本信息】 东南大学 ， 交通运输工程， 2021， 博士

【摘要】 交通运输是国民经济中战略性、引领性、基础性产业和服务性行业,建设“交通强国”是未来我国交通运输发展的总目标。随着移动互联网时代的开启,每个用户都成为了交通信息的贡献者,用户使用智能手机来规划路线、在线叫车、搜索目的地等。大量的基于位置的数据由这些设备和应用程序每天生成,包括在线订单,轨迹信息、地图查询数据和带地理标记的签到数据等。这些超大规模的多源数据在云端进行处理和融合生成城市全时段,无盲区的交通信息。面对如此庞大的数据,需要通过不断升级、完善与创新背后的智能交通决策系统,用机器学习和深度学习方法分析这些数据,助力智慧出行,从而减轻交通拥堵,促进城市道路管理水平提升。本文在深入分析现有的各种大规模交通预测任务的基础上,提出了“个性化智能交通推荐系统”这一概念,在智能交通系统中引入“个性化推荐”的思想,用以设计更加智能的交通系统。针对上述分析,本文依托国家自然科学基金优秀青年基金《多模式交通网络优化与管理》(71822007)和东南大学优秀博士学位论文培育基金(YBPY1927),以地理位置推荐、交通模式推荐、网约车调度单元推荐为研究背景,利用深度学习、强化学习等技术,为个性化智能交通推荐系统的设计提供方法和理论依据。本文的主要研究内容可以分成以下几个方面:首先,现有的推荐方法的核心思想都是围绕着用户行为进行,因此研究了用户行为的表示方法。针对出行行为大数据,提出基于用户行为的图嵌入学习方法,通过将用户的个性化出行行为编码在一个连续的向量空间中,进而作为监督学习模型的输入,以提高有监督学习模型的性能。其次,除了从个体角度进行推荐系统设计外,对于网约车调度单元推荐问题,需要从宏观角度进行交通状态的分析,对打车需求、司机分布等进行全局感知。针对时空数据,设计一种面向时空数据的交通状态预测方法,包括两个不同的注意力块,以捕获个性化的空间信息和时间信息,充分挖掘交通数据潜在的时空模式。并且从物理解释的角度解释了该模型,并探讨了神经网络架构对时空数据预测精度的影响。第三,传统的推荐算法有一定的局限性,这种局限性表现在可能会给用户大量同质化的推荐结果,造成这种问题的根源是没有考虑用户的偏好,也就是个性化的用户行为,这个问题在智能交通的应用上尤为突出。本文指出对出行者目的地的预测在本质上是一种地理位置的推荐,受推荐系统的启发,设计了一个两阶段的目的地预测框架,根据用户的历史行为生成个性化的候选集,这可以认为是粗排序,然后再借助特征工程对候选目的地进行精细的排序。第四,准确的交通模式推荐可以促进智能交通系统的发展,有助于缩短出行时间、缓解交通拥堵等。在多模式交通出行方式推荐问题中,不同用户在不同背景下对交通模式的偏好是不同的。本文针对交通模式推荐的应用场景,从用户、交通模式、地理位置和时间等多个角度进行系统的特征工程设计。为了更好的学习数据中存在的共现性,我们分别对历史数据中的O-D对和用户-OD对构建了一个二部图,然后利用图嵌入技术将图中的节点转化为特征向量。针对本研究存在的评价指标与损失函数不一致这个问题,我们提出了一种后处理算法,来处理预测结果与评价指标之间的不一致。最后,如何在日益增长的需求和有限的供给之间保持平衡,是网约车平台运营的核心问题。本文对网约车调度单元推荐的本质进行了探讨,指出了其本质上是负载均衡问题。借鉴推荐系统和负载均衡的思想,我们设计了算法的整体框架,强化学习返回排序后的推荐动作列表,然后以轮询形式将调度请求和动作进行匹配。通过合理的调度单元推荐,实现供需平衡,从而进一步减少乘客的等待时间,增加司机单位时间内的收益。更多还原

【Abstract】 Transportation is a strategic,leading,basic and service industry in the national economy.Building “China’s strength in transportation” is the general goal of China’s transportation development in the future.With the opening of the mobile Internet era,every user has become a contributor to traffic information.Users use smart phones to plan routes,call cars online,search for destinations,etc.A large amount of location-based data is generated by these devices and applications every day,including online orders,trajectory information,map query data,and geo-tagged check-in data.These super large-scale multi-source data are processed and fused in the cloud to generate the whole time traffic information of the city without blind spots.In the face of such a huge amount of data,it is necessary to continuously upgrade,improve and innovate the intelligent transportation decision-making system,analyze these data with machine learning and deep learning methods to help intelligent travel,thereby reducing traffic congestion and improving the level of urban road management.In this paper,based on the in-depth analysis of the existing large-scale traffic forecasting tasks,the concept of "Personalized Intelligent Transportation Recommendation System" is proposed.The idea of "personalized recommendation" is introduced into the intelligent transportation system to design a more intelligent transportation system.In view of the above analysis,this paper relies on the National Science Fund of China for Excellent Young Scholars "Multi-modal Transportation Network Optimization and Management"(71822007)and Southeast University Excellent Doctoral Dissertation Cultivation Fund(YBPY1927),and takes the geographical location recommendation,multi-mode travel mode recommendation and online car-hailing dispatching unit recommendation as the research background,using deep learning,reinforcement learning and other technologies to provide method and theoretical basis for the design of personalized intelligent transportation recommendation system.The main research contents of this paper are divided into the following aspects:First of all,the core ideas of the existing recommendation methods are all around user behavior,so the representation method of user behavior is studied.Aiming at the big data of travel behavior,a graph embedding learning method based on user behavior is proposed.By encoding the user’s personalized travel behavior in a continuous vector space,it is used as the input of the supervised learning model to improve the performance of the supervised learning model.Secondly,in addition to the design of the recommendation system from an individual perspective,for the recommendation of online car-hailing dispatching units,it is necessary to analyze the traffic status from a macro perspective,and make a global perception of taxi demand and driver distribution.Aiming at spatio-temporal data,a traffic state prediction method oriented to spatio-temporal data is designed,including two different attention blocks to capture personalized spatial and temporal information,and fully explore the potential spatio-temporal patterns of traffic data.The model is explained from the perspective of physical interpretation,and the influence of neural network architecture on the prediction accuracy of spatio-temporal data is discussed.Thirdly,the traditional recommendation algorithms have certain limitations,which may give users a large number of homogenized recommendation results.The root cause of this problem is that users’ preferences are not considered,that is,personalized user behavior.This problem is particularly prominent in the application of intelligent transportation.This paper points out that the prediction of traveler’s destination is essentially a recommendation of geographic location.Inspired by the recommendation system,a two-stage destination prediction framework is designed.A personalized candidate set is generated based on the user’s historical behavior,which can be regarded as rough sorting,and then the candidate destinations are sorted by feature engineering.Fourthly,accurate recommendations for multi-modal transportation modes can promote the development of intelligent transportation systems,help shorten travel time and alleviate traffic congestion.In the problem of multi-modal transportation travel mode recommendation,different users have different preferences for travel modes under different backgrounds.In this paper,according to the application scenarios of multi-modal travel mode recommendation,the system feature engineering design is carried out from the perspectives of users,travel mode,geographical location and time.In order to better study the co-occurrence in data,we construct a bipartite graph for O-D pairs and user-OD pairs in historical data,and then use graph embedding technology to convert the nodes in the graph into feature vectors.In view of the inconsistency between the evaluation index and the loss function in this study,we propose a post-processing algorithm to deal with the inconsistency between the prediction result and the evaluation index.Lastly,how to maintain a balance between increasing demand and limited supply is the core issue for the operation of online car-hailing platforms.In this paper,the essence of the recommendation of the online car-hailing dispatching unit is discussed,and it is pointed out that it is essentially a load balancing problem.Referring to the idea of recommendation system and load balancing,we design the overall framework of the algorithm,reinforcement learning returns the sorted recommendation action list,and then matches the scheduling request and action in the form of Round-Robin.Through reasonable dispatching unit recommendation,the balance between supply and demand can be achieved,thereby further reducing the waiting time of passengers and increasing the driver’s revenue per unit time.更多还原