摘要
针对高维数据近似最近邻查询,在过滤-验证框架下提出了一种基于学习的数据相关的c-近似最近邻查询算法.证明了数据经过随机投影之后,满足语义哈希技术所需的熵最大化准则.把经过随机投影的二进制数据作为数据的类标号,训练一组分类器用来预测查询的类标号.在此基础上计算查询与数据集中数据对象的海明距离.最后,在过滤后的候选数据集上计算查询的最近邻.与现有方法相比,该方法对空间需求更小,编码长度更短,效率更高.模拟数据集和真实数据集上的实验结果表明,该方法不仅能够提高查询效率,而且方便调控在查询质量和查询处理时间方面的平衡问题.
Under the filter-and-refine framework and based on the learning techniques, a data-aware method for c-approximate nearest neighbor query for high-dimensional data is proposed in this paper. The study claims that data after random projection satisfies the entropy-maximizing criterion which is needed by the semantic hashing. The binary codes after random projection are treated as the labels, and a group of classifiers are trained, which are used for predicting the binary code for the query. The data objects are selected who＇s Hamming distances between the query satisfying the threshold as the candidates. The real distances are evaluated on the candidate subset and the smallest one is returned. Experimental results on the synthetic datasets and the real datasets show that this method outperforms the existing work with shorter binary code, in addition, the performance and the result quality can be easily tuned.
出处
《软件学报》
EI
CSCD
北大核心
2012年第8期2018-2031,共14页
Journal of Software
基金
国家自然科学基金(60925008
60903014)
国家重点基础研究发展计划(973)(2010CB328106)
"核心电子器件
高端通用芯片及基础软件产品"国家科技重大专项(2010ZX01042-002-003-004)
关键词
随机投影
c-近似最近邻查询
支持向量机分类器
高维数据
熵最大化准则
位置敏感哈希
random projection
c-approximate nearest neighbor query
SVM classifier
high-dimensional data
entropy maximizing criterion
locality sensitive hashing
