III: Medium: Learning from Implicit Feedback Through Online Experimentation

该奖项是根据2009年《美国复苏和再投资法》(公法111-5)资助的。该项目的目标是利用用户与信息系统交互(例如，查询重构、点击、停留时间)中包含的信息，以培训这些系统更好地满足其用户的信息需求。关键的挑战在于正确解释这种隐含的反馈，并以一种提供有效训练数据的方式收集它。该项目超越了现有的被动数据收集方法，利用多臂强盗算法、实验设计和机器学习来主动收集隐式反馈数据。开发这些交互式实验方法的同时，还开发了可以使用所产生的训练数据的机器学习算法，以及验证算法假设的用户行为模型的经验评估。这项研究将提高企业内部网搜索和桌面搜索等重要应用的检索质量。此外，该项目将为物理学电子出版物Arxiv和可能的其他数字图书馆提供一个可操作的全文搜索引擎，从而形成研究的试验台，同时也向计算机科学以外的学术界提供宝贵的服务和传播工具。该项目为本科生和国际交换生提供了有趣和鼓舞人心的研究机会，PIS将把相关材料纳入本科生和研究生课程。最后，在他们之前在支持向量机、支持向量机(http://svmlight.joachims.org/))和用于学习排名检索功能和评估学习排名的开源搜索引擎之后，OSMOT(http://radlinski.org/osmot/)，)PI将继续提供易于使用的软件，通过项目网站(http://www.cs.cornell.edu/People/tj/implicit/).实现研究和教学

This award is funded under the American Recovery and Reinvestment Act of 2009 (Public Law 111-5).The goal of the project is to harness the information contained in users' interactions with information systems (e.g., query reformulations, clicks, dwell time) to train those systems to better serve their users' information needs. The key challenge lies in properly interpreting this implicit feedback and collecting it in a way that provides valid training data. Moving beyond existing passive data collection methods, the project draws on multi-armed bandit algorithms, experiment design, and machine learning to actively collect implicit feedback data. Developing these interactive experimentation methods goes hand-in-hand with developing machine learning algorithms that can use the resulting training data, and empirical evaluations that validate the models of user behavior assumed by the algorithms. This research will improve retrieval quality for important applications like intranet search and desktop search. Additionally, the project will provide an operational full-text search engine for the Physics E-Print ArXiv and potentially other digital libraries, thus forming a test-bed for the research while also providing a valuable service and dissemination tool to the academic community beyond computer science. The project provides interesting and motivating research opportunities to undergrads and international exchange students, and the PIs will include relevant material into the undergraduate and graduate curriculum. Finally, following their prior work on the Support Vector Machine, SVM-light (http://svmlight.joachims.org/) and an open-source search engine for learning ranked retrieval functions and evaluating the learned rankings, OSMOT (http://radlinski.org/osmot/), the PIs will continue to provide easy-to-use software that enables research and teaching, via the project website (http://www.cs.cornell.edu/People/tj/implicit/).