以数据仓库应用为代表的交互式查询分析技术为智能决策提供了支持。随着数据规模的不断增大,准确计算聚合查询结果往往需要全局数据扫描,使得这类查询面临着实时响应能力不足的问题。基于预先抽取的样本数据,复杂聚合查询提供快速的近...以数据仓库应用为代表的交互式查询分析技术为智能决策提供了支持。随着数据规模的不断增大,准确计算聚合查询结果往往需要全局数据扫描,使得这类查询面临着实时响应能力不足的问题。基于预先抽取的样本数据,复杂聚合查询提供快速的近似答案,在许多场景下是解决该问题的可行方案。分析了分层抽样优于随机抽样的具体条件,提出了一种两阶段分层抽样方法。首先针对业务特征进行分组,每个分组中使用随机抽样方法进行随机抽样,并评估其抽样效果。再针对抽样效果较差的分组,利用自组织特征映射网络(Self-organizing feature mapping,SOM)对数值进行聚类分组,改进其近似查询效果。基于公开数据集和实际电网数据的实验结果表明:本文方法相比于随机抽样、分层随机抽样以及国会抽样算法在相同抽样率下可达到15%的性能提升;与使用K-means、基于密度的聚类算法(Density-based spatial clustering of applications with noise,DBSCAN)等聚类方法相比,自SOM具有较好的近似查询结果。展开更多
Data uncertainty widely exists in many web applications, financial applications and sensor networks. Ranking queries that return a number of tuples with maximal ranking scores are important in the field of database ma...Data uncertainty widely exists in many web applications, financial applications and sensor networks. Ranking queries that return a number of tuples with maximal ranking scores are important in the field of database management. Most existing work focuses on proposing static solutions for various ranking semantics over uncertain data. Our focus is to handle continuous ranking queries on uncertain data streams： testing each new tuple to output highly-ranked tuples. The main challenge comes from not only the fact that the possible world space will grow exponentially when new tuples arrive, but also the requirement for low space- and time- complexity to adapt to the streaming environments. This paper aims at handling continuous ranking queries on uncertain data streams. We first study how to handle this issue exactly, then we propose a novel method （exponential sampling） to estimate the expected rank of a tuple with high quality. Analysis in theory and detailed experimental reports evaluate the proposed methods.展开更多
文摘以数据仓库应用为代表的交互式查询分析技术为智能决策提供了支持。随着数据规模的不断增大,准确计算聚合查询结果往往需要全局数据扫描,使得这类查询面临着实时响应能力不足的问题。基于预先抽取的样本数据,复杂聚合查询提供快速的近似答案,在许多场景下是解决该问题的可行方案。分析了分层抽样优于随机抽样的具体条件,提出了一种两阶段分层抽样方法。首先针对业务特征进行分组,每个分组中使用随机抽样方法进行随机抽样,并评估其抽样效果。再针对抽样效果较差的分组,利用自组织特征映射网络(Self-organizing feature mapping,SOM)对数值进行聚类分组,改进其近似查询效果。基于公开数据集和实际电网数据的实验结果表明:本文方法相比于随机抽样、分层随机抽样以及国会抽样算法在相同抽样率下可达到15%的性能提升;与使用K-means、基于密度的聚类算法(Density-based spatial clustering of applications with noise,DBSCAN)等聚类方法相比,自SOM具有较好的近似查询结果。
文摘Data uncertainty widely exists in many web applications, financial applications and sensor networks. Ranking queries that return a number of tuples with maximal ranking scores are important in the field of database management. Most existing work focuses on proposing static solutions for various ranking semantics over uncertain data. Our focus is to handle continuous ranking queries on uncertain data streams： testing each new tuple to output highly-ranked tuples. The main challenge comes from not only the fact that the possible world space will grow exponentially when new tuples arrive, but also the requirement for low space- and time- complexity to adapt to the streaming environments. This paper aims at handling continuous ranking queries on uncertain data streams. We first study how to handle this issue exactly, then we propose a novel method （exponential sampling） to estimate the expected rank of a tuple with high quality. Analysis in theory and detailed experimental reports evaluate the proposed methods.
