CAREER: Privacy-aware Predictive Modeling of Dynamic Human Events

利用个人事件数据的机器学习可以提高对未来事件的预测准确性，但会给每个人的隐私带来高风险。如今，大量的人类事件数据，如在线电视观看记录、域名服务器查询和医院入院的电子记录，越来越多地用于各种应用程序，包括网络分析和服务以及医疗保健分析。对这些集体事件序列进行预测建模，有利于促进全国经济和安全发展。例如，在网络流量诊断中，通过对用户活动的分析，可以预测和控制动态流量需求，提高风险响应效率。在健康信息学中，对患者入院事件的分析可以发现并优化处于风险中的个体的治疗，从而增强公共卫生准备和医疗保健结果。然而，通过优化单一的准确性目标，在历史事件数据上训练的机器学习算法可能会放大隐私风险。研究表明，可以从在线浏览历史和地点登记事件等人类活动中推断出人口统计和位置等私人属性。该项目旨在开发一个基于信任的机器学习框架，以更好地保护人类隐私，同时将对预测动态事件的效用的影响降到最低。机器学习和隐私的跨学科主题的研究和教育被整合到课程开发、学生研究项目和学术研讨会中。该项目开发了一系列新颖的模型和算法，在三个协同研究重点中分析动态的人类事件。(1)除了带有时间戳的事件序列，还可以利用事件类型、标签等附加的标记信息更好地捕捉事件之间的依赖关系。该项目研究了新的点过程、多视图学习和深度学习方法，用于分析具有事件标记信息的动态人类事件。(2)为了提高人类对预测建模的理解和信任，该项目开发了可解释的算法，以解释他们的信息如何用于事件预测，以及根据他们的输入可以推断出哪些潜在的私人信息。(3)平衡隐私和效用对个人和服务提供者都是有利的。该项目研究了用于事件预测的特定于用户的隐私保护方法，并通过将其表述为最小-最大优化问题来解决效用-隐私权衡问题。这三个研究目标的补充是在一些应用领域的综合评价。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。

Machine learning that leverages individuals' event data can improve the prediction accuracy of future events, but introduces high risks to each individual's privacy. Nowadays, large volumes of human event data, such as online TV-viewing records, domain name server queries, and electronic records of hospital admissions, are becoming increasingly available in a wide variety of applications including network analysis and services and healthcare analytics. Predictive modeling of those collective event sequences is beneficial for promoting nationwide economic and safety development. For example, in network traffic diagnosis, the analysis of user activities can be used to predict and control dynamic traffic demand, which improves risk response efficiency. In health informatics, the analysis of patient admission events can detect and optimize treatment for individuals at risks, which enhances public health preparedness and healthcare outcomes. However, by optimizing for the unitary goal of accuracy, machine learning algorithms trained on historic event data may amplify privacy risks. Studies have demonstrated that it is possible to infer private attributes such as demographics and locations from human activities such as online browsing histories and location check-in events. This project is to develop a trusting-based machine learning framework that better protects human privacy while minimally impacting utility for predicting dynamic events. Research and education on interdisciplinary topics of machine learning and privacy are integrated in curriculum development, student research projects, and academic seminars.The project develops a series of novel models and algorithms to analyze dynamic human events in three synergistic research thrusts. (1) Besides time-stamped event sequences, additional marker information such as event types and tags can be utilized to better capture the dependencies between events. This project investigates novel point processes, multi-view learning, and deep learning methods for analyzing dynamic human events with event marker information. (2) To improve human understanding and trust of predictive modeling, the project develops interpretable algorithms to explain how their information is used in event prediction and what potential private information can be inferred based on their inputs. (3) Balancing between privacy and utility is of mutual benefit to both individuals and service providers. This project investigates a user-specific privacy-preserving approach for event prediction and addresses utility-privacy tradeoff by formulating it as a min-max optimization problem. These three research aims are complemented by a comprehensive evaluation in a number of application domains.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.