• <object id="wmooc"><u id="wmooc"></u></object>
  • 首页> 外文学位 >文献详情
    【6h】

    Learning from time series in the resource-limited situations.

    机译:在资源有限的情况下从时间序列中学习。

    获取原文
    获取原文并翻译 | 示例
    ¥30

    摘要

    Many data mining tasks are based on datasets containing sequential characteristics, such as web search queries, medical monitoring data, motion capture records, and astronomical observations. In these and many other applications, a time series is a concise yet expressive representation. A wealth of current data mining research on time series is focused on providing exact solutions in such small datasets. However, advances in storage techniques and the increasing ubiquity of distributed systems make realistic time series datasets orders of magnitude larger than the size that most of those solutions can handle due to computational resource limitations. On the other hand, proposed approximate solutions such as dimensionality reduction and sampling suffer from two drawbacks: they do not adapt to available computational resources and they often require complicated parameter tuning to produce high quality results.;In this dissertation, we discuss anytime/anyspace algorithms as a way to address these issues. Anytime/anyspace algorithms (after a small amount of setup time/space) are algorithms that always have a best-so-far answer available. The quality of these answers improves as more computational time/space is provided. We show that by framing problems as anytime/ anyspace algorithms, we can extract the most benefit from the available computational resources and provide high-quality approximate solutions accordingly.;We further argue that it is not always effective and efficient to rely on whole datasets. When the data is noisy, using distinguishing local features rather than global features can mitigate the effect of noise. Moreover, building a concise model based on local features makes the computational time and space much less expensive. We introduce a new time series primitive, time series shapelets, as a distinguishing feature. Informally, shapelets are time series subsequences which are in some sense maximally representative of a class. As we shall show with extensive empirical evaluations in diverse domains, classification algorithms based on the time series shapelet primitives can be interpretable, more accurate, and significantly faster than state-of-the-art classifiers.
    机译:许多数据挖掘任务基于包含顺序特征的数据集,例如网络搜索查询,医疗监视数据,运动捕捉记录和天文观测。在这些应用程序和许多其他应用程序中,时间序列是一种简洁而又富有表现力的表示形式。当前大量的时间序列数据挖掘研究都致力于在如此小的数据集中提供精确的解决方案。但是,存储技术的进步和分布式系统的日益普及使现实的时间序列数据集由于计算资源的限制,比大多数这些解决方案可以处理的大小大几个数量级。另一方面,提出的近似解决方案如降维和采样有两个缺点:它们不适应可用的计算资源,并且它们通常需要复杂的参数调整以产生高质量的结果。算法来解决这些问题。任何时间/任何空间算法(经过少量的设置时间/空间之后)都是始终具有最佳答案的算法。随着提供更多的计算时间/空间,这些答案的质量也会提高。我们证明,通过将问题构造为随时随地的算法,我们可以从可用的计算资源中提取最大的收益,并据此提供高质量的近似解。;我们进一步认为依赖整个数据集并不总是有效和高效的。当数据嘈杂时,使用区分局部特征而不是全局特征可以减轻噪声的影响。此外,基于局部特征构建简洁的模型使计算时间和空间的成本大大降低。我们引入了一个新的时间序列原语,时间序列shapelet,作为一个显着特征。非正式地,小波是时间序列子序列,在某种意义上最大程度地代表了一个类。正如我们将在不同领域进行广泛的经验评估所表明的那样,基于时间序列shapelet原语的分类算法比最新的分类器可以解释,更准确并且显着更快。

    著录项

    • 作者

      Ye, Lexiang.;

    • 作者单位

      University of California, Riverside.;

    • 授予单位 University of California, Riverside.;
    • 学科 Computer Science.
    • 学位 Ph.D.
    • 年度 2010
    • 页码 172 p.
    • 总页数 172
    • 原文格式 PDF
    • 正文语种 eng
    • 中图分类 ;
    • 原文服务方 国家工程技术数字图书馆
    • 关键词

      ;

    相似文献

    • 外文文献
    站内服务

    联系方式:18141920177 (微信同号)

    客服邮箱:kefu@zhangqiaokeyan.com

    京公网安备:11010802029741号 ICP备案号:京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司?版权所有
    • 客服微信

    • 服务号

    婷婷五月亚洲综合图区_五十路A片熟妇在线观看_国产成人免费资源在线播放_欧美成人熟妇激情视频