毕业生知识库

城市出行轨迹挖掘对于理解个体出行行为、探究城市运行机理具有重要的意义。迈入大数据时代，随着移动设备、GPS终端等城市传感器的井喷式发展，更大时空范围内的城市活动轨迹得以精细化收集，为城市出行轨迹挖掘任务提供了全方位的数据支撑。然而，该任务存在着诸多挑战：首先，出行轨迹体量庞大，缺乏出行语义信息标注；其次，常规机理模型难以模拟复杂出行活动，而利用真实轨迹进行模拟会产生隐私泄露等问题；最后，以数据为支撑的场景模拟常存在宏、微观涌现结果冲突现象。平行交通系统能够有效应对这些挑战，它引入了与实际交通系统等价的人工交通系统，通过二者之间的虚实互动来实现对实际交通系统的管理与控制。本文面向平行交通系统，以手机信令、GPS轨迹为数据支撑，围绕出行语义感知、出行轨迹生成、场景辅助决策三个阶段循序渐进地开展对城市居民出行活动的挖掘。本文的研究内容与贡献如下：

1. 作为出行语义感知的研究内容之一，出行方式识别对于出行服务推荐等应用具有重要的意义。常规“分段-推理”的两阶段方法利用启发式规则识别出行方式切换点，不仅效率低而且分段准确性无法保证，同时，无标签轨迹中的可用信息难以被有效利用。本文对常规的两阶段方法进行改进，将出行方式识别任务重新建模为端到端的密集分类问题，并提出了基于相似熵正则化的半监督模型用于逐点标签推测。所提模型将带有快捷连接结构的编解码器网络作为骨干，以融合轨迹的底层物理特征与高层语义特征进行高精度识别。进一步地，本文设计了基于相似熵正则化的半监督模块以挖掘无标签轨迹中的隐含信息。在GeoLife轨迹数据集上的实验结果表明，所提模型的识别精度达到79.3%，并相较于基线模型在召回率指标上取得了5%左右的提升。

2. 作为出行语义感知的研究内容之二，推断居民的出行意图对于了解城市出行现状具有重要的意义。当前可用的数据中缺少相关语义标签，导致监督模型无法被有效利用。此外，出行相关的时空模式、地理信息等属性缺少统一的量纲，难以映射到同一特征空间中，导致出行模式提取困难。本文提出了一种协同经验过滤与主题模型的无监督出行意图推断方法，首先设计经验规则推测出行意图为工作和回家的行程，随后提出了一种连续变量感知的隐狄利赫雷分布模型，将用户的出行链视作文档、每次出行视为一个由异质属性组成的高维单词，以软聚类的方式识别用户剩余行程的出行意图。在北京市手机信令数据上的实验表明，所提模型相较于基线模型在困惑度指标中取得了7.7%的提升，且推断的结果与家庭出行调查的结果在宏观上具有高度的一致性。

3. 出行轨迹生成能够有效解决实际数据覆盖场景有限以及隐私泄露等问题。当前轨迹生成模型忽略了空间特性及访问位置时序性，导致生成结果与实际相差较大，此外，语义信息的缺失限制了生成结果的进一步应用。本文提出了一种语义信息指导的生成对抗网络，具体地，该网络将出行方式和出行意图两种语义项作为指导信号，利用基于掩码的多头注意力模块融合语义信息，以序列到序列的方式生成轨迹中的下一访问位置。为了令生成器学习到实际轨迹中隐含的空间模式，本文采用蒙特卡洛搜索将不完整的生成轨迹序列补全并转换为图像，随后用判别器来评估生成结果的真实性。在GPS轨迹数据上的实验表明，相比基线模型，所提模型将生成轨迹与真实轨迹在回旋半径和出行距离两种空间特性的差异分别降低了10%和27%，说明所提模型的生成结果更为真实。

4. 在平行交通系统的“实验与评估”运行模式下，本文以新冠肺炎疫情突发场景为切入点开展场景辅助决策研究。本文建模了病毒传播的个体动力学过程，并结合基于智能体的微观模型构建了人工社会交通系统。随后将出行轨迹加载到智能体中作为其交互依据，从微观视角揭示不同疫情防控措施下宏观群体的感染变化趋势。与现有研究工作不同，基于轨迹数据驱动的模拟无需个体间复杂拓扑网络等信息，通过判断出行轨迹的交汇即可确定智能体间的交互情况。此外，本文利用生成轨迹对不同人口密度、封控管理策略与疫情传播之间的关系进行了量化分析，充分展示了轨迹数据在场景决策中的价值。

综上所述，本文在平行交通系统的框架下开展城市出行轨迹挖掘的研究，提高了出行语义感知及出行轨迹生成等任务的准确性，在以数据驱动方式构建的人工交通系统中设计计算实验，从宏、微观层面探究了城市出行的传播动力学机制。研究成果有助于理解城市居民出行活动，充分发挥数据对智慧城市生产效率提高的乘数作用，对于理解城市行为、优化城市决策意义重大。

Mining urban mobility trajectory is critical for understanding individual travel behavior and exploring the mechanism of city evolution. In the era of big data, with the proliferation of ubiquitous mobile sensors such as cellphones and GPS devices, a wide range of spatio-temporal urban trajectories can be collected with fine resolutions, which provides comprehensive data support for the task of urban mobility trajectory mining. However, there are still many challenges in this task. First, the urban trajectories are with large volumes and their related semantic annotations are absent. Second, it is difficult to achieve impressive results when simulating complex human activities by mathematical models, while the use of real trajectories for simulation may involve risks such as privacy leakage. Finally, macroscopic and microscopic simulation results often conflict with each other. The parallel transportation systems (PTS) can effectively tackle these challenges by introducing artificial transportation systems that correspond to the real transportation systems. Through the virtual-reality interactions, PTS can realize the management and control for the real transportation systems. Oriented to the PTS, this dissertation carries out the mining of urban mobility trajectory step by step around three stages: travel semantic information awareness, travel trajectory generation, and scenario-assisted decision-making. The contents and contributions of this dissertation are as follows.

1. As one of the research components of travel semantic information awareness, transportation mode detection is important for applications such as transportation service recommendations. The conventional "segmentation-inference" two-stage pipeline uses heuristic rules to identify the change points in trajectories, which is not only inefficient but also cannot ensure the reliability of segmentation. In addition, the available information in unlabeled trajectories is difficult to be utilized effectively. This dissertation improves the conventional two-stage pipeline and re-models the transportation mode detection task as an end-to-end dense classification problem. Then, a semi-supervised end-to-end model with similarity entropy-based regularization is presented for point-wise detection. The proposed model applies an encoder-decoder network with shortcut connections as a backbone to fuse the shallow motion features with deep semantic features for high-precision detection. Further, a semi-supervised module based on similarity entropy regularization is designed to mine the underlying information in unlabeled trajectories. The experimental results on the GeoLife trajectory dataset show that the proposed model achieves 79.3% accuracy and about 5% improvement in the recall metric compared to baseline models.

2. As the second research component of travel semantic information awareness, inferring the travel intentions of residents is significant for understanding the current situation of city movements. The lack of related semantic labels in the available data leads to the inability of supervised models. In addition, there is a lack of unified measures for travel-related attributes, such as spatio-temporal patterns and geographic information, and it is thus difficult to map them into the same feature space, leading to inefficiency in travel pattern extraction. This dissertation proposes a collaborative travel intention inference approach based on empirical rules and a topic model. First, empirical rules are designed to infer and then filter out users' trips with job and homing intentions. Then, a continuous term aware latent Dirichlet allocation model is presented, in which the trip chains of each user and each trip are treated as a document and a high-dimensional word with heterogeneous attributes, respectively. The travel intentions for the remaining trips can be inferred in such a soft clustering way. Experiments on Beijing cellphone signaling data show that the proposed model achieves 7.7% improvement in the perplexity metric compared to the baseline models, and the inferred results are highly consistent with the results from the household travel survey at a macro level.

    3. The travel trajectory generation can effectively solve the problems of limited real-scenario coverage and privacy leakage. Existing trajectory generation models ignore the spatial characteristics and the temporal order of the visited locations, which leads to a large difference between the generated trajectories and the real ones. In addition, the lack of semantic information limits the further applications of the generated results. This dissertation proposes a semantic-guiding adversarial network for urban trajectory generation, in which the transportation modes and travel intentions are set as semantic items. The proposed model consists of masked multi-headed attention modules and yields the visits in a sequence-to-sequence manner. To learn the spatial characteristics implied by real trajectories, this dissertation uses a Monte Carlo search to complete the unfinished trajectories and converts them into images. Then a discriminator is adopted to signify how "real" the trajectory image looks, and its output is regarded as a reward signal to update the generator. The experiments on GPS trajectory data show that, compared to the baseline models, the proposed model can reduce the difference between the generated trajectories and real ones in two spatial characteristics (i.e., the radius of gyration and travel distance) by 10% and 27%, respectively, which indicates that the generated results are more realistic.

4. Under the "experimental evaluation" operation mode in PTS, this dissertation takes the epidemic outbreak scenario as a demonstration to perform the scenario-assisted decision-making task. The dynamic of virus propagation among individuals is modeled, and then an agent-based microsimulation environment is constructed as an artificial transportation system to conduct the epidemic spread simulation. After that, the trajectories are loaded into the agents as the basis for their interactions, and the macroscopic changes under different preventative measures can be revealed from the microscopic individual level. Unlike existing studies, there are no other requirements such as the complex topological networks of individuals in the trajectory data-driven simulation, in which the interactions between agents can be derived by the intersections of their trajectories. Additionally, this dissertation applies generated trajectories to explore the correlations of population densities and preventative measures with the epidemic spreading process. All of those demonstrate the decision-making value of urban trajectories.

To sum up, this dissertation studies urban mobility trajectory mining under the framework of the parallel transportation systems and improves the accuracy of tasks such as travel semantic information awareness and travel trajectory generation. By conducting computational experiments in a data-driven artificial transportation system, the propagation dynamics mechanism of urban travel is explored from macro and micro views. The achievements of the study help to understand urban mobility and bring into full play the multiplier effect of data on the productivity of smart cities, which is significant for understanding urban behavior and optimizing urban decision-making.