ATD: Collaborative Research: Real-Time Network Pattern Change Detection

来自互联网的快速增长的社交网络数据量为理解人类行为提供了大量信息。首先，网络数据包含稀疏的通信频率和一些密集的聚类，并且聚类随时间变化，因此特征的生成和选择至关重要。本研究项目解决了检测网络中突然类别变化的统计挑战。这对于量化人类动态和准确识别异常事件以及预测这些事件所表明的未来威胁非常重要。研究生将参与该项目的某些方面。该项目旨在开发1)静态情况：我们将使用零膨胀或障碍模型来表征类链接概率。2)对于动态情况：类通信概率是一个随时间变化的变量，采用自激过程对概率进行建模。3)我们考虑了冷启动问题，其中预测网络与训练网络相差很大，导致没有足够的样本来训练分类模型。相反，我们将开发矩阵变量聚类和分类模型。该项目包括几个重要的主题，以改进网络用户类别的建模，实时有效地识别网络模式的突变，以及减少异常值的影响。这些方法适用于各种类型的网络数据，如社会网络、生物信号、基因组序列等。pi将提供一个公开可用的软件包来实施建议的方法。此外，相应的统计理论和计算技术可以扩展到进一步的研究，并可以应用到其他领域。这个项目的主题是为了迎合学生在中佛罗里达大学新的大数据分析项目中进行实践研究。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The rapidly booming amounts of social networks data from the Internet offers a lot of information to understand human behaviors. First, the networks data contains sparse communication frequencies and some dense clusters, and the clusters change over time, so that feature generation and selection are essential. This research project addresses the statistical challenges for detecting abrupt categories changes in networks. This is important for quantifying human dynamics and accurately identifying unusual events and forecast future threats indicated by those events. Graduate students will be involved in some aspects of the project.This project aims to develop 1) for the static case: we will use zero-inflated or hurdle models to characterize the class link probability. 2) for the dynamic case: the class communication probability is a variable of time, we model the probability by a self-exciting process. 3) we consider the cold-start problem in which the predicted networks vary a lot from the training network, so that there are no enough samples to train classification models. Instead, we will develop matrix-variate clustering and classification models. This project includes several important topics to improve modeling of the network users' categories and identifying efficiently abrupt network pattern changes in real time as well as reducing the influence of outliers. These methods are applicable to various types of networks data such as social networks, biology signals, genome sequences, and so on. The PIs will provide a publicly-available software packages to implement the proposed methods. Additionally, corresponding statistical theories and computational techniques can be extended to advance further research and can be applied to other fields. This project topics cater to the students with hands-on studies in new Big-Data analysis program at the University of Central Florida.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.