III: Medium: Machine Learning with Humans in the Loop

III：媒介：人类参与的机器学习

• 批准号: 1513692
• 负责人: Thorsten Joachims
• 金额: $ 100万
• 依托单位: Cornell University
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2015
• 资助国家: 美国
• 起止时间: 2015-08-01 至 2020-07-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1513692&HistoricalAwards=false
• 关键词: III; Medium; Machine; Learning; Humans

Machine learning is increasingly used in systems common in daily life -- from search engines to online education to smart homes. Through interactions with their users, these systems learn about the world to improve efficiency and performance. Building effective and robust Human-Interactive Learning (HIL) systems, however, requires a framework that simultaneously integrates models of human behavior with the design of machine learning algorithms, because user data is only an indirect mapping, mediated through the human decision-making process, of the knowledge the system aims to elicit. This project takes an interdisciplinary approach to developing effective techniques to design human-interactive learning systems.This project explores how humans provide data and how learning algorithms use this data in an integrated framework that encompasses three aspects. First, the project develops models of the user decision process, formalizing how observable user actions map to the underlying knowledge the system aims to acquire. Second, these models inform the design of the interface that connects the user and the learning algorithm to suitably trade off the quantity and quality of the data acquired. Third, the user model and interface motivate new machine learning settings and algorithms to maximize learning efficiency. By developing an integrated framework for the three interconnected components for building Human Interactive Learning Systems -- human decision models, information-elicitation interfaces, and learning algorithms -- this project will impact future designs of widely-used systems such as non-web information search, recommendation, and online education.

机器学习越来越多地用于日常生活中常见的系统 - 从搜索引擎到在线教育到智能家居。通过与用户的互动，这些系统了解了世界以提高效率和性能。但是，建立有效且鲁棒的人际关系学习（HIL）系统需要一个框架，该框架同时将人类行为模型与机器学习算法的设计集成在一起，因为用户数据只是通过人类决策过程介导的间接映射，该系统的知识旨在激发该系统。 该项目采用跨学科的方法来开发有效的技术来设计人际关系学习系统。该项目探讨了人类如何提供数据以及学习算法如何在包含三个方面的集成框架中使用此数据。首先，该项目开发了用户决策过程的模型，并形式化了可观察的用户行动如何映射到系统旨在获取的基础知识。其次，这些模型为连接用户和学习算法的接口设计提供了信息，以适当地权衡所获取数据的数量和质量。第三，用户模型和接口激发了新的机器学习设置和算法，以最大程度地提高学习效率。 通过为建立人类互动学习系统的三个互连组件开发一个集成框架 - 人类决策模型，信息引起的接口和学习算法 - 该项目将影响广泛使用的系统的未来设计，例如非WEB信息搜索，建议搜索，建议和在线教育。

项目成果

The Impact of More Transparent Interfaces on Behavior in Personalized Recommendation

• DOI: 10.1145/3397271.3401117
• 发表时间: 2020-07
• 期刊: Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval
• 作者: Tobias Schnabel;Saleema Amershi;Paul N. Bennett;P. Bailey;T. Joachims