REU Site: University of North Carolina at Greensboro in Complex Data Analysis using Statistical and Machine Learning Tools

在这个大数据时代，技术创新允许以低成本收集大量数据。数百GB、TB甚至PB的大小已不再罕见。适当处理和分析这些数据是对未来一代毕业生的重要培训。考虑到这一目标，拟议的REU计划旨在提供为期10周的“使用统计和机器学习工具进行复杂数据分析”的复杂培训，以在2020-2022年夏季期间为八（8）名来自数学科学的高度积极性的国家选拔本科生提供培训。教师导师带来了丰富多样的经验，这个培训计划。 PI Gupta，Co-PI Gao，高级人事Richter和Stufken是统计学家;高级人事Mohanty是计算机科学学院，专门研究机器学习工具;高级人事Sun是数学和统计学系的统计遗传学家。该培训计划将激励学生参与者，特别是那些来自代表性不足的少数民族的学生，继续攻读数学科学研究生课程，并成为能够处理社会数据分析需求的训练有素的专业人员。 作为更广泛的专业培训的一部分，学生将前往北卡罗来纳州的主要研究中心，如SAS，SAMSI（统计和应用数学科学研究所），以及纳米科学和纳米工程联合学院。我们希望作为本次培训的一部分完成的研究将是非常高质量的，并将导致期刊文章和会议报告。 数据的复杂性可以通过多种方式来实现。数据的高维度是这样一种复杂性，其中变量的数量与数据大小相比可能相对较大。数据污染是另一种类型的复杂性。学生将学习同时处理降维和异常值检测的艺术，作为项目之一的一部分。添加噪声的数据（在公开发布之前创建数据的机密性）是另一种类型的复杂性。对于研究人员来说，只能访问加密数据而不能访问真实的数据的情况越来越普遍。在其中一个项目中，我们将讨论为什么以及如何使用随机响应模型对数据进行加密和解密，保留聚合级别属性并确保受访者的匿名性。REU计划的机器学习部分将专注于利用复杂工具的功能，如无监督和监督机器学习和深度学习，以识别与疾病症状相关的社交媒体帖子，并预测疾病传播的时间趋势。违反模型假设是另一个复杂性的来源，需要使用非参数技术，如reservation方法。在许多配对设计情况下，完整和不完整的数据对的混合是可用的。而不是忽略不完整的数据对，我们将培养学生的方法设计用于分析这些数据。在另一个项目中，学生将接受子数据选择技术的培训，这对处理巨大规模的数据非常有帮助。这个项目将解决以下类型的问题：（1）子数据应该有多大才能确保可靠的分析;（2）对于给定的大小，应该如何选择子数据？所有这些项目的首要目标是培养学生认识数据中的各种复杂性，并找到正确的技术来处理这些数据。该奖项反映了NSF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。

In this age of big data, technological innovations allow the collection of massive amounts of data at low cost. Sizes of hundreds of gigabytes, terabytes, and even petabytes are no longer uncommon. Appropriate handling and analysis of such data is vital training for the future generation of graduates. With this goal in mind, the proposed REU program aims to provide 10-week sophisticated training in “Complex Data Analysis using Statistical and Machine Learning Tools” to eight (8) highly motivated nationally selected undergraduates from Mathematical Sciences during summers of 2020-2022. The faculty mentors bring a rich and diverse experience to this training program. PI Gupta, Co-PI Gao, Senior Personnel Richter and Stufken are statisticians; Senior Personnel Mohanty is a Computer Science faculty specializing in machine learning tools; and Senior Personnel Sun is a statistical geneticist in the Department of Mathematics and Statistics. The training program will motivate the student participants, particularly those from under-represented minorities, to go on to graduate programs in mathematical sciences and become better trained professionals capable of handling societal data analytics needs. As part of broader professional training, students will undertake trips to major research centers in North Carolina such as SAS, SAMSI (Statistical and Applied Mathematical Sciences Institute), and the Joint School of Nano Science and Nano Engineering. We expect that the research completed as part of this training will be of very high quality and will lead to journal articles and conference presentations. Complexity in data can come in a variety of ways. High dimensionality of the data is one such complexity where the number of variables can be relatively large as compared to the data size. Data contamination is another type of complexity. Students will learn the art of simultaneous handling of dimensionality-reduction and outlier detection as part of one of the projects. Noise-added data (to create confidentiality in data before public release) is another type of complexity. It is becoming common for researchers to have access only to scrambled data, and not the real data. In one of the projects, we will talk about why and how data are scrambled and de-scrambled using randomized response models, retaining aggregate level properties and ensuring anonymity to respondents. The machine learning component of the REU program will focus on leveraging the capabilities of sophisticated tools such as Unsupervised and Supervised Machine Learning, and Deep Learning for identification of social media posts related to disease symptoms, and for prediction of temporal trends in disease propagation. Violation of model assumptions is another source of complexity that necessitates the use of nonparametric techniques such as the resampling methods. In many matched-pairs design situations, a mixture of complete and incomplete pairs of data are available. Rather than ignoring data from incomplete pairs, we will train students in methods designed for analyzing such data. In another project, students will be trained in the subdata selection techniques which are very helpful in dealing with data of enormous size. This project will address questions of the type (1) what size should the subdata have to ensure a reliable analysis; and (2) for a given size, how should the subdata be selected? The overarching goal in all of these projects will be to train students in recognizing various complexities in data and finding the right techniques to handle such data.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.