ICES: Small: Economic Analysis of Recommender Systems

推荐系统获取在线用户历史和人口统计信息，并自动创建针对商品和服务的个性化推荐。这种系统在书籍和电影等小物品中已经很常见，随着新的应用出现，比如公司寻找供应商和工人寻找工作，这种系统注定会在经济中发挥越来越大的作用。然而，以前的研究几乎没有阐明推荐系统带来的经济利益和成本，也没有关于多种系统中的哪一种在当前或未来的利基市场中工作得最好。该项目将改进衡量推荐系统性能的指标，使用这些指标评估主要利基市场的现有推荐算法，并开发和测试适用于服务不足的利基市场的新算法和系统。这项研究建立在计算机科学家之前的研究基础上，并将其与既定的经济理论以及人类受试者的受控实验联系起来。改进后的指标明确包含了用户、供应商和平台提供商的经济目标。这些算法包括协作过滤、基于内容的过滤和混合算法，以及更高级的算法，这些算法使用不同的信息，并在与用户交互的同时主动权衡探索和利用。这些数据将包括已经为该项目编制的专有行业数据，以及在加州大学圣迭戈分校进行的受控实验室实验和在线现场实验中获得的新数据。直接影响将是增加对各种推荐系统的优点和缺点的了解。更广泛地说，这项研究将阐明人类如何处理免费获得的信息，这是许多社会科学学科中的一个重要问题，而推荐系统的数百万用户最终可能会获得实际好处。此外，开展这项研究将促进担任研究助理和程序员的研究生、为编程做出贡献的本科生(主要是游戏设计工程专业)以及数百名将作为实验室受试者的本科生的在校教学、培训和学习。

A recommender system takes on-line user histories and demographic information, and automatically creates personalized recommendations for goods and services. Already common for small items such as books and movies, such systems are destined to play an increasing role in the economy as new applications emerge for them, such as firms searching for suppliers and workers searching for jobs. Previous research, however, has shed little light on the economic benefits and costs that recommender systems bring, nor on which of the many sorts of systems work best in each current or prospective niche.This project will improve metrics for measuring performance of recommender systems, use these metrics to evaluate existing recommender algorithms in major niches, and develop and test new algorithms and systems adapted to under-served niches. The research builds on previous research by computer scientists, and connects it to established economic theory as well as to controlled experiments with human subjects. The improved metrics explicitly incorporate the economic goals of users, suppliers, and platform providers. The algorithms include collaborative filtering, content based filtering, and hybrids as well as more advanced algorithms that use heterogeneous information and actively trade off exploration and exploitation while interacting with the users. The data will include proprietary industry data already compiled for this project, together with new data obtained in controlled laboratory experiments conducted at UCSC and in on-line field experiments.The direct impact will be to increase understanding of the strengths and weaknesses of various sorts of recommender systems. More broadly, the research will shed light on how humans process freely available information, an important question in many social science disciplines, and practical benefits may ultimately accrue to millions of users of recommender systems. Additionally, conducting this research will promote on-campus teaching, training and learning for graduate students who serve as research assistants and programmers; for undergraduates (mostly game design engineering majors) who contribute to the programming; and for hundreds of undergraduates who will serve as laboratory subjects.