ATD: Collaborative Research: Point Process Algorithms for Threat Detection from Heterogeneous Human Mobility and Activity Data

该项目旨在开发新的算法和模型，用于分析人类生成的、时空标记的事件数据，以便快速识别可能代表活动威胁或即将发生的威胁的异常模式。这项工作背后的一般激励原则是，这类事件，特别是与人类流动性密切相关的事件，通常表现出强大的模式，因此，无论是在个人还是集体层面上，与典型模式的重大偏离都是值得怀疑的，应该被视为潜在的危险。因此，该项目将使当局能够更好地快速确定何时出现威胁情况，以便他们能够迅速做出反应，减轻潜在的破坏性后果。为了实现这一总体目标，将实现几个子目标：1)构建模型，以整合不同粒度级别的人类事件的不同数据集，从基于个人到基于邻域，再到基于区域；2)开发一个框架，将这些整合的数据与人类流动性模型联系在一起；3)创建方法，在给定的框架内，在不同的尺度上，包括个人、个体群体或空间区域，检测数据的突然变化。该项目将构建一个框架，用于分析来自不同来源的人类时空事件数据，该框架以随机点过程的数学为基础，但专门为表示其底层结构与人类流动性密切相关的事件而量身定做。几个想法将在这个框架中统一起来，这样产生的算法就能够识别可能代表持续或新出现的威胁的异常行为或事件。首先，将探索新的方法对高频人体流动性数据进行预处理--例如GPS轨迹数据--以降低维度并更好地适应标记点过程框架。下一步，将开发新的标记点过程类别，它们能够更好地处理通常隐藏在空间人类事件数据下的详细几何结构，因为这类事件通常是在人类运动的规律性基础上分层的；针对这一任务，专门提出了几何嵌入人类流动性主题的高阶Hawkes过程。为了更好地识别可能表现出异常行为或对异常行为作出反应的地理区域或个人子集，将开发新的方法，在不同的抽象和物理相关性水平上对受这些点过程影响的数据进行聚集，从个人到相关的社会群体，再到社区。为此，将开发新的方法来快速检测数据中的异常，与预期的点过程相比，同样是在不同的聚类级别上，通过拟合优度衡量；这里提出了一种贝叶斯方法，使用转移学习来检测即使是非常有限的数据集的新模式。最终的结果是一套单独有用的工具，这些工具的组合将成为一种强大的新方法，可以组织和分析人类事件的大型数据集，以检测威胁行为。

This project aims to develop new algorithms and models for analyzing human-generated, space-time marked event data in order to quickly identify anomalous patterns that may represent active threats or threats about to occur. The general motivating principle behind the work is that such events, especially when intimately tied to human mobility, generally display robust patterning, such that significant deviations from the typical patterns, either on an individual or collective level, are cause for suspicion and should be treated as potentially dangerous. Thus, this project will better enable authorities to quickly determine when threatening situations arise, so that they can react to them rapidly, mitigating potentially devastating consequences. To achieve this overarching objective, several sub-objectives will be met: 1) constructing models to integrate disparate datasets on human events at varying levels of granularity, from individual-based to neighborhood-based to region-based; 2) developing a framework that ties this integrated data together with models of human mobility; 3) creating methods for detecting sudden changes within the data given the framework, at varying scales, including individuals, groups of individuals, or spatial regions. This project will construct a framework for analyzing human spatio-temporal event data arising from heterogeneous sources, based on the mathematics of stochastic point processes, but specifically tailored to represent events whose underlying structure is intimately tied to human mobility. Several ideas will be united in this framework, such that the resulting algorithms are able to identify anomalous behavior or events that may represent ongoing or emerging threats. First, new methods will be explored for pre-processing high frequency human mobility data -- eg., gps trace data -- to reduce dimensionality and better fit within the marked point process framework. Next, new classes of marked point processes will be developed that are better able to handle the detailed geometric structure often underlying spatial human event data, given the regularity of human motion upon which such events are often layered; high-order Hawkes processes geometrically embedding human mobility motifs are proposed specifically for this task. New methods will be developed for clustering data subject to these point processes at varying levels of abstraction and physical relevance, from individuals, to linked social groups, to neighborhoods, in order to better identify geographic regions or subsets of individuals that may be displaying or reacting to anomalous behavior. To go along with this, new ways will be developed to quickly detect anomalies within the data as compared to the expected point process through goodness of fit measures, again at differing levels of clustering; proposed here is a Bayesian method to detect emerging patterns for even very limited datasets using transfer learning. The end result with be a suite of tools that are all individually useful, and that combined will serve as a powerful new method of organizing and analyzing large datasets of human events to detect threatening behavior.