查询结果:   邬志罡,荆一楠,何震瀛,王晓阳.基于用户查询与样本间匹配度评估的分层抽样策略[J].计算机应用与软件,2019,36(8):196 - 202.
中文标题
基于用户查询与样本间匹配度评估的分层抽样策略
发表栏目
算法
摘要点击数
703
英文标题
A STRATIFIED SAMPLING APPROACH BASED ON MATCHING DEGREE EVALUATION BETWEEN USER QUERY AND SAMPLE SET
作 者
邬志罡 荆一楠 何震瀛 王晓阳 Wu Zhigang Jing Yi’nan He Zhenying Wang Xiaoyang
作者单位
复旦大学计算机科学技术学院 上海 201203 上海市数据科学重点实验室(复旦大学) 上海 200433 上海智能电子与系统研究院 上海 200433   
英文单位
School of Computer Science, Fudan University, Shanghai 201203, China Shanghai Key Laboratory of Data Science, Fudan University, Shanghai 200433, China Shanghai Institute of Intelligent Electronics and Systems, Shanghai 200433, China   
关键词
抽样系统 近似查询处理 分层抽样 优化问题
Keywords
Sampling system Approximate query processing Stratified sampling Optimization problem
基金项目
国家自然科学基金项目(61732004);国家重点研发计划项目(2018YFB1004404);上海科技创新行动计划项目(16DZ11002001)
作者资料
邬志罡,硕士生,主研领域:数据管理。荆一楠,讲师。何震瀛,副教授。王晓阳,教授。 。
文章摘要
在数据探索性分析场景下,用户倾向于借助抽样系统获取近似查询结果来换取更快的查询速度。现有的抽样系统通常假设用户的历史查询记录能很好地表征未来的查询情况,从而针对特定的查询特征生成特定的抽样策略。然而,在现实场景中,用户探索意图变化丰富,用户查询特征的稳定性假设通常无法得到保证。为解决上述问题,提出一种评估任意用户查询与样本间匹配度的方法。离线训练生成多份样本集,并在应对具体查询时自动选取最匹配样本集进行近似结果计算。离线样本集的生成是以在所有可能的用户查询上的预期匹配度损失总和最小作为训练目标。实验结果表明,在真实数据集上,该抽样系统与现有方法相比,将近似结果的精确度提高了26.3%。
Abstract
During the data exploration tasks, users usually prefer to use sampling system for getting an approximate answer rather than suffer from high query latency. Existing sampling systems usually make hypothesis that the historical user query workload can represent the pattern of future user queries very closely. Based on this hypothesis, they specifically design sampling strategy for specific user query pattern. However, in the real use case, the users’ exploration intentions are always changing, so the hypothesis of the stability of the user query pattern cannot be guaranteed. To solve these problems, this paper proposed a method to evaluate the matching degree between any user query and the sample set. The system generated multiple offline sample sets. When a particular user query came, the system could automatically choose the best matching sample set and calculate the approximate query answer. The offline sample sets were trained so that the expected total sum of the matching degree losses upon all possible user queries became the lowest. The experimental results show that, compared with the existing methods, the accuracy of the approximate results is improved by 26.3% on the real data set.
下载PDF全文