III: Small: Entity Selection and Ranking for Data-Mining Applications

像linkedin.com、odesk.com和guru.com这样的专家管理门户网站是指导性的网站，允许人们向更广泛的公众宣传他们的工作或技能。例如，linkedin拥有超过1.2亿的会员，这使得潜在的雇主、合作者等能够发现具有所需专业知识的个人或个人群体。类似地，像亚马逊或Yelp这样的评论管理网站收集了大量关于产品或服务的评论。例如，kindle在亚马逊上有3万多条评论。当然，用户不可能遍历所有这些评论，并且可以通过识别一小部分足够有信息量的评论得到很大的帮助。最后，随着在线社交和媒体网络作为新闻和其他信息来源的重要性日益提高，迫切需要能够自动识别和推荐网络重要节点的工具，特定用户可能需要关注这些节点以充分利用在线社交媒体的力量。在这些场景中，给定一组实体（例如，关于产品的评论、宣布某些技能的专家、网络节点或边缘），目标是确定重要实体的子集（例如，有用的评论、称职的专家、有影响力的节点）。关于推荐系统的现有工作试图通过实体排名或实体选择来识别重要实体。实体排名方法将分数与每个实体关联起来；它们忽略了高分实体之间的冗余。实体选择方法试图通过评估一组实体的合意性来克服这一缺点；它们试图识别实体的最佳子集，而忽略其他可能与最佳子集一样好或几乎一样好的实体子集。在此背景下，本项目旨在克服现有实体选择和实体排名方法的缺点，通过将两者协同集成到一个共同的框架中，允许基于实体选择的实体排名和基于实体排名的实体选择。在最终的框架中，单个实体的得分部分取决于它们可以属于的良好实体群体的数量；好的实体组是由得分高的实体组成的。这项工作解决的主要挑战是如何探索组合问题的解空间，以识别参与许多好的解的实体子集。由此产生的探索组合问题解空间的新实用方法发现了与专家管理系统、在线产品评论管理和网络分析（包括物理和社会网络）相关的应用。该项目还为波士顿大学的研究生和本科生提供了更多的研究型培训机会。所有的研究成果，包括出版物、软件和数据将通过项目网站：http://www.cs.bu.edu/~evimaria/sel-and-ranking.html免费传播给更广泛的研究和教育界。

Expert-management portals like linkedin.com, odesk.com and guru.com are indicative sites that allow people to advertise their work or set of skills to the broader public. For example, linkedin features more than 120 million members which allows potential employers, collaborators, etc. to discover individuals or groups of individuals with the desired expertise. Similarly, review-management sites like Amazon or Yelp collect large number of reviews about products or services. For example, kindle has more than 30,000 reviews on Amazon. Naturally, users cannot go over all these reviews and are helped significantly by the identification of a small subset of reviews that is sufficiently informative. Finally, as online social and media networks grow in importance as sources of news and other information, there is an urgent need for tools that automatically identify and recommend important nodes of the network, that specific users may need to follow to fully exploit the power of online social media. In each of these scenarios, given a collection of entities (e.g., reviews about a product, experts that declare certain skills, network nodes or edges), the goal is to identify a subset of important entities (e.g., useful reviews, competent experts, influential nodes respectively). Existing work on recommender systems attempts to identify important entities either by entity ranking or by entity selection. Entity-ranking methods associate a a score with each entity; They ignore the redundancy between the highly-scored entities. Entity-selection methods try to overcome this drawback by evaluating the desirability of a group of entities taken together; They attempt to identify the best subset of entities, while ignoring other subsets of entities that may be equally-good or almost as good as the best subset. Against this background, this project aims to overcome the drawbacks of existing entity selection and entity ranking methods through a synergistic integration of both into a common framework that allows entity-ranking based on entity selection and entity-selection that based on entity ranking. In the resulting framework, the scores of individual entities are determined in part by the number of good groups of entities they can be part of; and good group of entities consist of entities with high scores. The main challenge addressed by this work is how to explore the solution space of combinatorial problems in order to identify subsets of entities that participate in many good solutions. The resulting new practical methods for exploring the solution space of combinatorial problems find applications related to expert management systems, management of online product reviews, and network analysis (including physical and social networks). The project also offers enhanced opportunities for research-based training of graduate and undergraduate students at Boston University. All of the research results including publications, software, and data will be freely disseminated to the broader research and educational community through the project website at: http://www.cs.bu.edu/~evimaria/sel-and-ranking.html.