行为数据驱动的个体间关系结构协同发现与匹配的模型与算法研究

Individual behavior modeling and behavior based relational structure inference is one of the fundamental methods for techniques such as behavior prediction, user profiling. It can also be applied in more concrete problems such as social marketing, preventive device maintenance. In fact, the relations among individuals are often hidden and it calls for effective models and methods to quantitatively infer the underlying relational graph structure behind the behavior and event data. This project is based on the previous research obtains on behavior modeling and graph matching. It mainly addresses the problems for relational graph inference and discovery based on behavioral event data, along with multiple graph matching and disambiguation originated from heterogeneous relations. The project mainly covers: (1) Based on stochastic point process, develop general frameworks and machine learning algorithms for the problem of discovering relational graph structure from behavioral event data; (2) Depending on specific applications, develop tailored models and algorithms by exploring the particular structure of the problem, with the purpose of addressing the challenge of high-dimension, large-scale, and heavy-noise during model learning. (3) Perform simultaneous matching for a set of the independently inferred graph structures, which allows the information to transfer over multiple graphs, with the aim of correcting the biased estimation due to inherent ambiguity from independent single graph structure inference. This project explores the behavioral event data-driven relational graph structure inference and matching, which provides some theoretical insights and practical algorithms for real-world applications.

个体行为事件的建模及基于行为数据的个体间关系推断是行为预测、用户画像分析等技术的核心方法之一，并可用于社会化营销、预防性设备维保等具体问题。事实上，个体间关系往往无法直接观测，需要发展相应的模型和算法从行为事件数据来定量推断行为背后的潜在关系图结构。本申请拟结合在行为建模和图匹配方面的前期积累，研究面向行为事件数据的关系图推断和发现，以及多个关系图间的协同匹配与消歧问题，主要包括：（1）基于行为事件数据，在随机点过程框架下，研究关系图推断的相关模型和算法；（2）面向具体应用场景，挖掘特定结构定制算法，应对实际问题维度高、规模大、噪声强的挑战；（3）对基于行为数据得到的关系图结构，进行跨多个图的协同匹配，形成图结构信息的传递机制，以此减小或者消除基于单个关系图上行为事件信息进行估计时的歧义性。本项目立足于行为数据驱动的关系图结构推断与匹配的模型设计，具有理论意义和应用前景。

结构信息是数据科学中的重要形式，特别是机器学习的一个重要研究对象。本项目不仅从传统图像等数据中提取图结构信息，也从其他常见的行为数据中挖掘图结构信息。本项目主要在以下几个方面展开了研究并取得了相关研究成果：1）面向图像数据，这些步骤包括图像中关键点的检测与提取，图结构的构建与学习，以及跨图的协同关键点图神经网络学习与匹配；2）在此基础上，面向行为数据，基于时序点过程、时序状态空间模型等技术，研究多维时序行为数据的潜在关系结构，实现高维时序数据的隐含图结构关系挖掘与推断；3）最后，面向例如社交网络等大规模网络数据，基于图嵌入、马尔科夫随机游走等技术，研究协同的关系图边链接预测与跨图匹配方法。本课题研究丰富了面向图与序列数据的协同建模和机器学习的理论和方法，取得了一定基础研究成果，并有助于推动这一领域成果的实际应用。