RI: Small: Collaborative Research: Statistical ranking theory without a canonical loss

RI：小：协作研究：没有典型损失的统计排名理论

• 批准号: 1320894
• 负责人: Pradeep Ravikumar
• 金额: $ 22.36万
• 依托单位: University of Texas at Austin
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2013
• 资助国家: 美国
• 起止时间: 2013-08-01 至 2017-08-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1320894&HistoricalAwards=false
• 关键词: RI; Small; Collaborative; Research; Statistical

The problem of ranking objects occupies a central place in key technologies such as web search and recommendation systems. These technologies have a tremendous daily impact on the lives of millions of people. Moreover, the enormous scale of data on the web makes the use of machine learning especially attractive in constructing ranking algorithms. A huge amount of research effort has been devoted to developing efficient ranking algorithms that can deal with a variety of data sets encountered in web search and recommendation systems.This project develops unifying mathematical theory that will provide a basis for understanding and categorizing existing algorithms and, more importantly, lead to deeper insights and new algorithms for the problem of learning to rank. The investigators also apply ranking algorithms to new domains. For example, ranking chemical reactions based on their plausibility will help chemists discover much-needed reaction bases for technologies such as carbon dioxide reduction, and conversion of natural gas into gasoline. Fundamental advances in the statistical theory of ranking will be incorporated into undergraduate and graduate courses. Data sets and software developed will be made freely available to the scientific community. The investigators will also organize a workshop with a focus on interdisciplinary participation and involvement of under-represented groups in computer science and statistics.The primary technical challenge in developing statistical ranking theory is the absence of a universally agreed-upon loss functions for ranking. This is in contrast to classic machine learning problems such as classification and regression, where there are only a few natural possibilities for the loss function and these are well-understood theoretically. The project addresses this gap by investigating how different loss functions for ranking affect fundamental theoretical properties such as learnability, and by creating a theory of convex surrogates that is applicable when loss functions abound. The project re-examines existing statistical literature on ranking with a computational lens. This will enable development of flexible and efficient plug-in decision rules that model the conditional probability of labels given inputs.By incorporating the results of this research into courses and survey articles, the PIs help train a new generation of machine learning researchers and practitioners who will view ranking as a learning problem on par with classification and regression in mathematical depth as well as practical importance. Theoretical guidance for practitioners formulating new algorithms for ranking will improve the most common applications on the web.

对象排序问题在网络搜索和推荐系统等关键技术中占据着核心地位。这些技术对数百万人的日常生活产生巨大影响。此外，网络上的海量数据使得机器学习在构建排名算法方面特别有吸引力。 大量的研究工作致力于开发有效的排名算法，可以处理网络搜索和推荐系统中遇到的各种数据集。该项目开发了统一的数学理论，将为理解和分类现有算法提供基础，更重要的是，为学习排名问题带来更深入的见解和新算法。研究人员还将排名算法应用于新领域。 例如，根据化学反应的合理性对化学反应进行排名将有助于化学家发现二氧化碳还原和将天然气转化为汽油等技术急需的反应基础。排名统计理论的基本进展将纳入本科生和研究生课程。开发的数据集和软件将免费提供给科学界。研究人员还将组织一个研讨会，重点关注计算机科学和统计学领域的跨学科参与和代表性不足的群体的参与。发展统计排名理论的主要技术挑战是缺乏普遍认可的排名损失函数。这与分类和回归等经典机器学习问题形成鲜明对比，在这些问题中，损失函数只有几种自然可能性，而且这些问题在理论上是可以很好理解的。该项目通过研究不同的排名损失函数如何影响可学习性等基本理论属性，并创建适用于损失函数大量存在的凸代理理论，来解决这一差距。 该项目从计算角度重新审视现有的排名统计文献。 这将有助于开发灵活高效的插件决策规则，对给定输入的标签的条件概率进行建模。通过将这项研究的结果纳入课程和调查文章中，PI 有助于培训新一代机器学习研究人员和从业者将把排名视为一个学习问题，在数学深度和实际重要性上与分类和回归同等。 对制定新排名算法的从业者的理论指导将改进网络上最常见的应用程序。