III: Medium: Collaborative Research: Optimization with Sparse Priors--Algorithms, Indices, and Economic Incentives

这是一个合作研究项目，结合了斯坦福大学的Ashish Goel(IIS-0904325)和宾夕法尼亚大学的Sanjeev Khanna(IIS-0904314)的专业知识。传统上，内容是由有限数量的出版商(如书店、音乐公司和报纸)生成的，然后由专业编辑和评论员对其质量进行评估。然而，近年来，个人已经成为内容的大规模生产者，以分散的方式产生图像、博客、意见和推荐。然后，这些内容被其他用户发现和使用，而由于可用内容的巨大规模，集中式审查变得不可行。因此，需要利用显式和隐式的用户反馈，以便向互联网用户提供最佳排名和推荐。同样广泛的问题也出现在在线广告、讨论板的自动审核以及社交网络上用户偏好的自动推断上。除了非常大之外，互联网上的用户活动数据通常也非常稀疏，因为每个用户只执行一小部分可能的动作(例如，搜索一小部分关键字、评论或购买一小部分产品)。该项目旨在设计算法和优化技术，以有效地利用这些数据。稀疏数据被视为对用户偏好的“先验信念”。该项目还旨在设计经济激励措施，以获得有用和正确的数据，对操纵具有健壮性。这项研究的两个部分相互作用很强，因为算法部分可以识别出需要获取的有价值的附加信息。这两个部分结合在一起，可以帮助用户从互联网数据中获得最佳价值。这个项目的结果将提高搜索引擎的性能，并促进使用用户反馈的网络应用程序。项目网站(http://www.stanford.edu/~ashishg/sparse_opt.html)将用于传播成果。

This is a collaborative research project combining the expertise of Ashish Goel, Stanford University (IIS-0904325) and Sanjeev Khanna, University of Pennsylvania (IIS-0904314). Traditionally, content has been generated by a limited number of publishers (such as book houses, music companies, and newspapers), and its quality then evaluated by professional editors and reviewers. In recent years, however, individuals have become mass producers of content, generating images, blogs, opinions, and recommendations, in a decentralized manner. This content is then discovered and consumed by other users, and centralized review is rendered infeasible by the sheer magnitude of available content. Consequently, there is a need to utilize user feedback, both explicit and implicit, in order to provide optimum rankings and recommendations to Internet users. The same broad problem occurs in online advertising, automatic moderation of discussion boards, and automated deductions of user preference on social networks. In addition to being very large, user activity data on the Internet is also typically very sparse, since each user only performs a small share of possible actions (e.g., searches for a small fraction of keywords, reviews or purchases a small fraction of products). This project aims to design algorithms and optimization techniques to effectively utilize such data. The sparse data is treated as a "prior belief" on user preferences. The project also aims to design economic incentives to obtain useful and corrective data, robust to manipulation. The two parts of this research interact strongly with each other, since the algorithmic component can identify valuable pieces of additional information to acquire. Together, these two parts can help users derive optimum value from Internet data. Results of this project will improve search engine performance and facilitate web applications that employ user feedback. The project Web site (http://www.stanford.edu/~ashishg/sparse_opt.html) will be used to disseminate results.