面向大规模异构数据的哈希表征学习研究

发布时间：2018-07-11 18:47

本文选题：哈希表征 + 异构数据　；参考：《清华大学》2016年博士论文

【摘要】：随着异构数据的快速增长,异构数据间的检索出现效率瓶颈。本文提出引入哈希这种简单的表征形式来加速异构数据间的检索。然而,传统哈希方法无法对异构数据的异构性建模,因此,无法应用于异构数据检索。本文针对异构数据的三个异构性(包括数据特征的异构性,数据对象的异构性以及数据网络的异构性)提出了相应的哈希学习框架,从而实现了异构数据上高效而精准的相似检索。本文的主要贡献如下:1.提出面向异构特征的哈希学习:针对属性特征高维稀疏及二值的特性和内容特征连续低维稠密的特性,本文提出了两个定制的概率生成模型分别对两种不同的特征建模,从而,充分地融合了两种特征的信息,使得哈希方法的精度得到显著提升。2.提出面向异构对象的哈希学习:本文提出了基于关联的异构哈希方法(Ra HH),实现了跨数据域的高效相似检索。Ra HH将不同数据域分别映射到不同的汉明空间,并基于异构关联关系学习汉明空间之间的映射。从而既保持了不同数据域的特性,又实现了跨数据域的检索。3.提出面向异构网络的哈希学习:本文主要关注关系不可传递的网络和有向网络。本文分别提出了基于相似成份的哈希学习方法及保持高阶相似度的哈希学习方法,对两种网络进行了合理建模,从而实现了异构网络数据上的高效相似检索。
[Abstract]:With the rapid growth of heterogeneous data, the efficiency of heterogeneous data retrieval appears bottleneck. In this paper, a simple representation of hash is proposed to accelerate the retrieval of heterogeneous data. However, the traditional hash method can not model the heterogeneity of heterogeneous data, so it can not be applied to heterogeneous data retrieval. In this paper, a hashing learning framework is proposed for the three heterogeneity of heterogeneous data (including the heterogeneity of data characteristics, the heterogeneity of data objects and the heterogeneity of data networks). Thus, the efficient and accurate similarity retrieval on heterogeneous data is realized. The main contributions of this paper are as follows: 1. Aiming at the high dimensional sparse and binary characteristics of attribute features and the continuous low dimensional dense characteristics of content features, two customized probability generation models are proposed to model the two different features respectively. The accuracy of the hashing method is greatly improved by fully integrating the information of the two features. A heterogeneous object oriented hash learning is proposed. In this paper, an association-based heterogeneous hash method (Ra HH) is proposed, which implements the efficient similarity retrieval across data domains. Ra HH maps different data domains to different hamming spaces. And learning the mapping between the Han Ming space based on the heterogeneous association relationship. Thus, it not only maintains the characteristics of different data domains, but also realizes the retrieval of. 3. 3 across data domains. Hashing learning for heterogeneous networks is proposed: this paper focuses on networks with intransisible relationships and directed networks. In this paper, hashing learning method based on similarity component and hashing learning method based on maintaining high order similarity are proposed, and two kinds of networks are modeled reasonably, so as to realize efficient similarity retrieval on heterogeneous network data.
【学位授予单位】：清华大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.3

【相似文献】