群智感知中基于社会属性的行为引导与激励研究

Fueled by the proliferation of sensor-enabled smart devices and rapid development of Mobile Internet, Crowdsensing is now emerging as a new method to perceive this physical world and complex society. Crowdsensing assigns collaborative sensing task through Mobile Internet and collects sensory data from massive smart devices carried by people. Numerous applications of Crowdsensing have been developed in recent years towards scheduling smart cities, traffic monitoring, indoor localization, community behavior analysis, and so on. More importantly, people, who carry smart devices such as smartphones and tablets, becomes the leading role of sensing task, which is traditionally undertaken by sensor networks. Inspired by this change, we study the methodology of steering behaviors and Incentive mechanisms through analyzing the social characteristics of participants, as well as the algorithms of heterogeneous data fusion. We first discovery the potential relations between social networks and the sensing task, and then construct dynamic evaluation model on users’ behaviors, based on which to design the policies of steering participants in crowdsourced sensing. Key factors determining the sensing effectiveness, e.g., spatial information and task attributes, are also quantified and modeled to improve the quality of data and to reduce sensing costs. By leveraging contextual information of participants, we propose a framework for directly increasing quality of sensing and formalizing the incentive mechanism. Expected results of this project should provide important references and support technologies, both theory and applications, for researchers and decision makers using Crowdsensing.

随着移动互联技术和智能终端计算能力的快速发展,群智感知成为复杂物理世界和社会感知的新方法，其利用大众智能设备上的各种传感器，并通过移动互联网络进行任务分工与协作，可广泛应用于智慧城市规划、交通状况实时监控和社群行为分析等诸多领域。由于感知任务的主体从传统的传感器网络过渡到了大量的携带有智能设备的人身上，因此感知任务主体的选择和协同方式成为主要面临的挑战。本项目以群智感知的主体作为切入点，研感知任务究参与者的社会属性，设计行为引导策略和任务激励机制，并研究异构数据的融合方法。首先分析社会属性特征与感知任务的关系；其次建立任务参与者的行为动态评估模型，设计参与者行为协调和引导策略；最后对社会属性特征值和位置信息等关键影响力因素进行量化统计和建模，并基于其上下文信息对感知数据进行推导和预测。本项目的预期研究成果将为设计和开发高效的群智感知计算平台提供可行的策略依据和共性支撑技术。

本项目主要聚焦于群智感知网络中的各种关系和模式的挖掘和学习，利用社交网络提供的信息（包括社交关系，用户位置轨迹信息，用户属性等），采用深度学习的方法进行数据挖掘，从而实现对用户的运动模式，身份，社交关系，兴趣点以及行为动态的理解、评估与推断。项目组的主要研究内容包括了三个大的方向：时空数据挖掘，社交网络数据挖掘及图元学习。在时空数据挖掘方面，基于用户轨迹数据，研究了轨迹-用户的链接问题，运动模式及用户社交关系推断问题等，相继提出了TULER、TULVAE、CATHI、Vec2Link、SCIMAC等一系列模型。在社交网络数据挖掘中，基于社交网络平台数据对POI推荐，话题推荐，人口统计属性推断，跨网络身份识别以及信息传播模式等进行了研究，前后提出了VASER、VANext、APOIR、DeepTrip、DeepAttr、CasCN、DMT-LIC、VaCas、NTP、DeepLink、dNAME等一系列模型。在图元学习方面，提出了图元学习中的少样本节点分类模型Meta-GNN，以及将图元学习方法应用于网络对齐提出了基于图元学习的快速网络对齐方法Meta-NA。通过大量理论和实验验证，与现有的最好模型进行对比，上述模型达到了在相关子领域里面的现有最好成绩，验证了本课题成果的先进性。