CAREER: New Paradigms for Online Machine Learning

我们生活在一个技术驱动的世界里，用户每天都在与大规模的自动化系统进行交互。在线推荐系统、搜索引擎和个性化医疗只是以机器学习（ML）算法为核心的系统的几个例子。机器学习作为一个领域的长期成功依赖于将其转化为一种易于使用、无缝的技术，并具有严格的、可证明的性能保证。此外，为了对社会产生积极影响，机器学习技术需要具备应对任何大型多用户系统所伴随的社会挑战的能力。这个CAREER项目的首要目标是使对社会负责的ML成为一种易于访问的黑盒技术，适用于大型多用户交互系统。该项目特别关注三个具体挑战。第一个挑战是通过自动化设计任务特定ML算法的过程，使ML成为即插即用技术。第二个挑战是为现代应用开发机器学习方法，比如预测社交网络中的用户偏好，因为社交网络中的数据是不断发展的，并且相互关联复杂。第三个挑战是开发对社会负责的推荐系统的理论和算法，并且不会使其用户两极分化。近年来，通过探索概率论与序列预测问题之间的内在联系，为在线学习提供了统一的理论和算法设计原则。这个CAREER项目将以这些发展为基础。利用概率论中所谓的伯克霍尔德方法和数学规划领域的进步，该项目将致力于自动化设计新的有效在线学习算法的过程。基于最近发展的在线放松的想法，该项目将引入新的方法来设计计算效率高的算法，用于从相互关联的数据点中学习。最后，使用和扩展经典统计学的思想来处理控制和讨厌的变量，该项目将为推荐系统开发新的方法，以避免用户两极分化。CAREER计划将通过开发与ML相关的新教育组件来推进STEM教育。“面向大众的机器学习”（Machine Learning for the mass）研讨会将与康奈尔大学（Cornell）的Women In Computing共同组织，旨在让女性和代表性不足的少数族裔参与进来，并让本科生在成长阶段接触机器学习领域的研究和工作机会。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。

We live in a technology-driven world, where users interact with large-scale, automated systems on a daily basis. Online recommendation systems, search engines and personalized medicine are just a few examples of systems that use Machine Learning (ML) algorithms at their core. The long term success of ML as a field relies on transforming it into an easily usable, seamless technology with rigorous, provable guarantees on performance. Further, to have a positive impact on society, ML technologies need to be equipped to handle the social challenges that accompany any large multi-user systems. The overarching goal of this CAREER project is to make socially-responsible ML a readily accessible black-box technology that is applicable in large multi-user interactive systems. In particular, the project focuses on three concrete challenges. The first challenge is to make ML a plug-and-play technology by automating the process of designing task specific ML algorithms. The second challenge is to develop ML methods for modern applications such as predicting user preferences in social networks, where data is evolving and complexly interconnected. The third challenge is to develop theory and algorithms for recommendation systems that are socially responsible and do not polarize its users. In recent years, exploring inherent connections between probability theory and sequential prediction problems have lead to a unifying theory and algorithm design principles for online learning. This CAREER project will build on these developments. Using the so called Burkholder method from probability theory and advances in the field of mathematical programming, the project will aim at automating the process of designing new and effective online learning algorithms. Building on the recently developed idea of online relaxations, the project will introduce novel methodology for designing computationally efficient algorithms for learning from interconnected data points. Finally, using and extending ideas from classical statistics to deal with control and nuisance variables, the project will develop new methods for recommender systems that can avoid polarizing users. The CAREER program will advance STEM education by developing new educational components related to ML. ``Machine Learning for the Masses'' workshops will be co-organized with Women In Computing at Cornell aimed at involving women and underrepresented minorities and exposing undergraduates to research and job opportunities in the field of ML during their formative years.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.