Data mining and OLAP over sequential data

研究人员正在研究日益庞大的分布式数据集。文学研究人员可以获取大多数数字格式的经典作品。万维网上有数兆字节的生物数据可以免费获得。文本或时间序列等时序数据很常见，已被数据库研究界确定为一个至关重要的领域，利用用户驱动的技术挖掘以TB为单位的时间序列或文本模式是一个重要的开放问题。我们希望为用户提供工具，以便他们可以交互地将顺序数据聚合到视图中，就像在联机分析处理(OLAP)中一样。为此，我们需要数据上经过良好调整的层次结构，以便能够支持向下钻取和向上滚动查询。我们需要稳健的时间序列索引。最后，我们需要一个序列数据库的良好模型：我们应该补充现有的信息检索工具还是扩展现有的数据库？在任何一种情况下，数据量都是如此之大，我们需要良好的策略来预聚合数据。在OLAP中，属性值是按星型架构层次结构(商店/城市/省份、客户ID/年龄组)聚合的。单词可以聚合成短语、句子和段落，也可以按下义关系(汽车/车辆)聚合，尽管词义消除歧义存在问题。时间序列的一个等价物是分段。分段将数据分成几个间隔，在这些间隔中，数据根据简单的模型(例如，恒定、线性、二次、单峰、单调、凸形)进行近似处理。分段是一个层次化的过程，因为间隔可以进一步细分。最近，时间扭曲索引取得了一些进展。当比较两个时间序列时，时间扭曲局部地加速或减慢一个时间序列中的时间，以计算稳健的相似性度量。

Researchers work over increasingly large distributed data sets. Literary researchers have access to most classical works in digital format. There are terabytes of biological data freely available on the Web. Sequential data such as text or time series is common and has been identified as an area of critical importance  by the database research community.Mining patterns in terabytes of time series or  text with user-driven techniques, is an important open problem. We want to provide tools to users so that they can interactively aggregate sequential data into views, as in On-Line Analytical Processing (OLAP). To do so, we need well adapted hierarchies over the data so that drill-down and roll-up queries can be supported. We need robust time series indexing. Finally, we need a good model for sequential databases: should we complement existing information retrieval tools or extend current databases? In either case, the amount of data is so large that we need good strategies to preaggregate the data.In  OLAP, attribute values  are aggregated by star schema hierarchies (store/city/province, client ID/age group). Words can be aggregated in phrases, sentences and paragraphs or by hyponymy (car/vehicle) though there are issues with word sense disambiguation. An equivalent for time series is segmentation. Segmentation divides the data into intervals where it behaves approximately according to a simple model (e.g. constant, linear, quadratic, unimodal, monotonic, convex). Segmentation is  a hierarchical process since intervals can be further subdivided.Recently, some progress has been made in time warping indexing.  When comparing two time series, time warping locally accelerates or slows time in one time series to compute robust similarity measures.