Introduction to special issue on recommender systems
Introduction to special issue on recommender systems
复制标题
DOI：
10.1145/1921591.1921592
复制
发表时间：
2011-02
期刊：
影响因子：
--
通讯作者：
J. Riedl;Barry Smyth
中科院分区：
文献类型：
--
作者：
J. Riedl;Barry Smyth
推荐系统正在改变人们与网络互动的方式。从亚马逊这样的电子商务网站到digg和slashdot这样的新闻和信息网站，推荐服务通过提供更加个性化的信息访问体验，帮助人们在多样化的产品和复杂的信息中进行选择。在本期特刊中，我们寻求确定成熟的研究，这些研究不仅关注基础和算法，还关注推荐系统技术在新兴网络技术领域的新应用。经过全面的同行评议，四篇文章入选。共提交了29篇文章，每篇由该研究领域的3位专家审阅。所提交文章的强烈回应和主题的多样化体现了当今推荐系统研究的活力。几乎每篇文章都有有趣而发人深省的见解，但只有四篇文章可以被接受为特刊。在Netflix挑战赛向获胜团队颁发了100万美元奖金之后，从业者承认，用于获奖的基本方法——将数百个单独的算法与一个强大的机器学习框架结合起来——由于缺乏可扩展性，无法在实践中使用。本期特刊的第一篇文章由Cacheda等人撰写，研究了当前协同过滤技术在高性能推荐系统的可扩展性需求背景下的局限性。特别是，他们使用流行的评估指标对一系列当代协同过滤算法进行了全面的比较评估，此外，他们还提出了一些新的指标，旨在强调某些重要的推荐属性。此外，作者提出了一种新颖的替代协同过滤技术，该技术侧重于用户和项目之间的差异，而不是它们的相似性，尽管这种方法很简单，但报告的结果具有竞争性，特别是在较大的数据集中。总的来说，推荐系统的成功很大程度上依赖于构成其用户模型或配置文件基础的偏好信息。例如，协作过滤风格的系统依赖于根据用户的显式评级或交易历史构建的基于评级的配置文件。这些数据通常是稀疏的，这最终限制了推荐的机会，并且在实践中，如果概要数据有限，许多系统都会面临冷启动问题。Umyarov和Tuzhilin的工作提出使用外部聚合评级信息作为处理此类挑战的一种可能方法，以提高推荐性能。特别是，它们描述了外部总体评级如何……
Recommender systems are changing the way people interact with the Web. From e-commerce sites like Amazon.com to news and information sites like digg and slashdot, recommenders help people choose among diverse products and complex information by providing a more personalized information access experience. For this special issue we seek to identify mature research that focuses not just on foundations and algorithms but also on novel applications of recommender systems technologies in emerging areas of Web technologies. Four articles have been selected as a result of a comprehensive peer-review process. In total 29 articles were submitted, each reviewed by 3 experts in this field of research. The remarkably strong response and the diverse topics of the submitted articles exemplifies the vitality of research in recommender systems today. Nearly every article had interesting and through-provoking insights, but only four could be accepted for the special issue. After the Netflix Challenge awarded one million dollars to the winning team, practitioners confessed that the fundamental approach used to win the prize—combining literally hundreds of individual algorithms with a powerful machine-learning framework—was impossible to use in practice because of its lack of scalability. The first article in this special issue, by Cacheda et al., examines the limitations of current collaborative filtering techniques in the context of the scalability demands of high-performance recommender systems. In particular, they provide a comprehensive comparative evaluation of a range of contemporary collaborative filtering algorithms using popular evaluation metrics and, in addition, propose a number of new metrics that seek to emphasize certain important recommendation properties. Moreover, the authors present a novel alternative collaborative filtering technique that focuses on the differences between users and items, rather than their similarities and, despite the simplicity of this approach, report competitive results especially in larger datasets. By and large, the success of a recommender system is critically dependent on the preference information that forms the basis of its user models or profiles. For example, collaborative filtering-style systems depend on ratings-based profiles constructed from the explicit ratings or transactional history of users. This data is often sparse, which ultimately limits recommendation opportunities, and, in practice, many systems are faced with the cold-start problem if profile data is limited. The work of Umyarov and Tuzhilin proposes the use of external aggregate rating information as one possible way of dealing with such challenges in order to improve recommendation performance. In particular, they describe how external aggregate ratings …