Bayesian Methods for Large-Scale Applications

摘要建议编号：DMS-0505599机构：罗格斯大学新不伦瑞克国家科学基金会项目：统计学家：Madigan，DavidPROPOSAL标题：大规模应用的贝叶斯方法研究人员致力于大规模应用的贝叶斯统计方法。有三个应用程序为这项工作提供了背景。“文本分类”涉及将文档自动分配到预先定义的类别，并需要超高维的有监督学习模型。“作者归属”使用类似的方法，但试图识别匿名文档的作者。“本地化”问题使用信号特征来定位无线网络中的用户。研究人员关注跨越这些应用的技术挑战，包括序贯贝叶斯分析、非线性优化和新颖性检测算法。在商业和科学领域，计算的进步已经极大地改变了数据分析的角色。从历史上看，分析人员在当地产生数据以解决具体的研究问题。现在，无处不在的计算和廉价的存储使数据的生产与研究问题脱钩。各种数据的产生和储存在可远程访问的数据库中，心中有无数的问题，既有可预见的，也有不可预见的。从历史上看，统计学的重点是从有限的数据中提取最大量的信息。相反，研究人员的工作重点是针对这些新兴的更大规模应用的所谓贝叶斯统计方法

ABSTRACT PROPOSAL NUMBER.: DMS-0505599 INSTITUTION: Rutgers University New Brunswick NSF PROGRAM: STATISTICSPRINCIPAL INVESTIGATOR: Madigan, DavidPROPOSAL TITLE: Bayesian Methods for Large-Scale Applications The investigators work on Bayesian statistical methods for large-scaleapplications. Three applications provide the backdrop for the work."Text categorization" concerns the automatic assignment of documentsto predefined categories and requires ultra-high dimensional supervisedlearning models. "Authorship attribution" uses similar methodologybut attempts to identify authors of anonymous documents.The "Localization" problem uses signal characteristics to locate users in wireless networks. The investigators focus on technical challenges that span these applications including sequential Bayesian analysis,non-linear optimization, and novelty detection algorithms.In both the business and scientific realms, computing advances havedrastically altered the role of data analysis. Historically, analystsproduced data locally to address specific research questions. Now,ubiquitous computing and cheap storage have decoupled the productionof data from the research questions. Data of all kinds are producedand deposited in remotely accessible databases with myriad questionsin mind, both foreseen and unforeseen. Statistics has historicallyfocused on squeezing the maximum amount of information out of limiteddata. The investigator's work focuses instead on so-calledBayesian statistical methods for these emerging larger-scale applications