CIF: Small: Collaborative Research: Network Event Detection with Multistream Observations

网络事件检测的目标是检测大型网络上是否存在一组节点，这些节点的观测结果反映了异常事件的发生。现有的网络事件检测研究主要从两个角度进行。第一种是数据驱动的，不假设任何底层统计模型，通常适用于更一般的数据集，但可能没有性能保证。第二种视角是模型驱动的，假设数据具有一定的统计分布（例如，高斯分布），通常具有性能保证，但可能仅限于数据符合模型的应用程序。这个项目的目标是探索一个网络事件检测框架，该框架统一了广泛的事件检测问题，其中数据被假设由一些潜在的统计分布控制，但在某种意义上是数据驱动的，很少有关于分布的先验假设。所开发的检测方法和统计工具具有广泛的应用，包括欺诈检测、临床试验、医疗诊断、高频交易、投票违规分析和网络入侵。在这个项目中，通过探索三个重点，开发了一种针对一般网络事件检测问题的综合方法：(i)（非结构化）点事件的检测，（ii）基于图的结构化事件的检测，以及（iii）动态演变图结构的顺序和最快检测。所设计的测试的性能是根据检测错误的概率和该错误趋于零的速率来表征的。解决了各种基本问题，包括非身份证。数据流，以及网络大小、事件大小、样本大小和数据维度之间的相互作用。

The goal in network event detection is to detect the existence of a set of nodes over a large network whose observations reflect the occurrence of an unusual event. Existing studies of network event detection have been mainly from two perspectives. The first is data-driven without assuming any underlying statistical model, and is typically applicable to more general data sets, but may not come with performance guarantees. The second perspective is model-driven, with certain statistical distributions (e.g., Gaussian) assumed for the data, and usually comes with performance guarantees, but may be limited to applications where the data fit the model. The goal in this project is to explore a framework for network event detection that unifies a wide range of event detection problems, in which the data are assumed to be governed by some underlying statistical distributions, but is data-driven in the sense that little is assumed a priori about the distributions. The developed detection approaches and statistical tools have a wide range of applications, including fraud detection, clinical trials, medical diagnosis, high-frequency trading, voting irregularity analysis, and network intrusion.A comprehensive approach to general network event detection problems is developed in this project through the exploration of three thrusts: (i) detection of (unstructured) point events, (ii) detection of graph-based structured events, and (iii) sequential and quickest detection of dynamically evolving graph structures. The performance of the designed tests is characterized in terms of the probability of detection error and the rate at which this error goes to zero. Various fundamental issues are addressed, including non-i.i.d. data streams, as well as the interplay between network size, event size, sample size, and data dimension.