【作者】 强保华；

【导师】 吴中福； 吴开贵；

【作者基本信息】 重庆大学 ， 计算机软件与理论， 2005， 博士

【摘要】 信息技术的迅猛发展和Internet 技术的广泛应用,给人们的生活、工作和学习带来了极大的方便。但随着人们对信息综合利用需求的进一步提高,Internet的局限性也凸现出来,因为计算机互连网络实现的是一种硬件的连通,所连接起来的信息大部分是异构的,形成了一个个信息孤岛,很难达到互联互通,真正实现数据的共享。信息的不一致造成有限的信息共享、缺少数据交换是制约实现数据综合利用的主要瓶颈。因此,如何找出并消除异构数据库中数据的冲突、异常,进行数据库集成,是实现综合信息应用必须解决的问题。数据库集成技术可以有效地检测数据异常、尽早地调整数据、消除数据的不完整和不一致,从而有助于提高信息综合利用的质量。异构数据库集成的关键问题是找出异构数据库间相同的语义对象,即语义集成,具体到关系数据库中,就是要找出异构数据库间相同的属性和实体(记录),即属性匹配和实体匹配问题。本篇论文主要在分析研究了现有异构数据库语义集成技术特点的基础上,针对目前方法的不足,考虑到神经网络具有较强的自学习能力和推广能力,比较适合处理语义集成这类不确定性问题,将神经网络理论应用到异构数据库语义集成中的属性匹配和实体匹配领域。同时,在解决实体匹配问题时,针对属性权重难以准确计算的问题,我们分别考虑了属性的信息熵和互信息来估算属性的权重。从提高识别相同语义对象的查准率和查全率两个重要指标出发,提出了新的属性匹配和实体匹配算法,并在具体应用环境中的数据库上进行了实验,取得了理想的效果。本论文主要研究工作如下: ①全面地综述了目前异构数据库集成所涉及的主要问题,详细地介绍了异构数据库语义集成的任务、语义异构的类型和解决语义异构问题主要采用的方法,并对目前方法的特点和不足之处进行了研究,分析了将神经网络理论、信息熵和互信息理论引入异构数据库语义集成领域中解决属性匹配和实体匹配问题的可行性。②分析了目前文献中基于BP 神经网络的属性匹配方法的不足,指出不同的输入在神经网络上可能对应相同的输出是影响神经网络结果准确性的主要因素之一,并进行了理论上的证明和具体实验环境的验证。为了解决该问题,在论文中提出了用相同的训练数据在不同的初始权值和阈值的情形下多次训练神经网络,从而构建不同于一般多级分类器方法的多个分类器的思想,可以实现对干扰数据的有效过滤。该思想在本文的第三章和第五章得到了很好的体现。更多还原

【Abstract】 With the great developments and extensive applications of information technology and internet, the life, work, and study of people become more convenient. But with the increasing requirements of comprehensive information usage, the defaults of internet are becoming obvious. Computer networks just connect the hardwares, and the data on different computers are heterogeneous and become information islands, on which data sharing and interoperability become more and more difficult. Limited data sharing, difficult data communications, and inconsistent formats are the main bottlenecks for realizing data comprehensive usage. So how to detect and resolve data conflicts and heterogeneity, and integrate the heterogeneous database to realize data sharing and comprehensive usage are the fundamental issues for new information technology application. Database integration technologies provide effective channels for detecting data heterogeneity, modifying data earlier, resolving data incompleteness and inconsistencies, and finally improving data quality for comprehensive usage. Finding the corresponding semantic objects, i.e. the semantic integration, is the most important issue in heterogeneous database integration domain. Exactly, the task for semantic integration is to find the corresponding attributes and entities in relational database domain. Based on the research background and the characteristics of the present heterogeneous database semantic integration techniques, the dissertation is intend to develop research on heterogeneous database semantic integration using neural network considering the self learning and generalization abilities of neural network. In the dissertation, identification of corresponding semantic objects, i.e. attributes matching and entity matching, is mainly studied. At the same time, in order to estimate the attribute weight accurately, attribute entropy and mutual information are also considered to be used to calculate attribute weight when solving entity matching problem. New algorithms for attribute matching and entity matching are proposed, and the experimental results show the algorithms can improve the precision and recall obviously. The main contributions of the dissertation are summarized as follows: ①The present main issues for heterogeneous database integration are surveyed comprehensively; the task of heterogeneous database semantic integration, the types of semantic heterogeneity, and the approaches to resolving semantic heterogeneity are introduced in detail; the existing problems of present semantic integration techniques are also studied. Finally the feasibility to resolve attribute and entity matching problems are analyzed using neural network, entropy, and mutual information. ②The defaults of attribute matching based on BP neural network in present literatures are analyzed. It is pointed out in the dissertation that different inputs for neural network may have the same outputs is the main factor to decrease the accuracy of neural network, which is illustrated and tested. So the idea for establishing multi classifiers is proposed, in which the neural networks are trained several times using the same training data set with different initialized connective weights and thresholds. Our proposed multi classifiers can filter the interference data effectively. The effectiveness of this approach is tested in chapter three and chapter five. ③Due to the existing problems of attribute matching in heterogeneous database semantic integration, two-phase-check algorithm for attribute matching based on BP neural network is presented according to the idea to establish multi classifiers, in which attributes are required to be categorized according to data types firstly, then the BP neural network architecture is defined according to the characteristics of categorized attributes, and the BP neural networks are trained several times respectively using the categorized attributes with different initialized connective weights and thresholds. In the procedure of attributes matching, the attribute characteristic vector is input into the corresponding network, and the final attribute matching result are the intersection of every time matching result in different neural networks. The experimental results show our proposed approach can improve the attribute matching accuracy and decrease the training time obviously. ④Considering the existing defaults of weights assignment on entity matching during heterogeneous database semantic integration, an approach to computing attributes weights based on attribute entropy and the decision model for entity matching are proposed in the dissertation as well as the matching algorithm for heterogeneous entities based on attribute entropy. Our proposed approach can make good use of the instance values information of attributes, which is objective and is easily quantified. The experimental results on real-world data indicate our approach can get high accuracy. Meanwhile, different attributes with the same entropy will get the same weights directly by computing the attribute entropy, which can not effectively differentiate the importance of attributes. So mutual information between attributes is considered and used to compute attribute weight as well as attribute entropy. In thedissertation, the algorithm for computing final entropy of attribute is also proposed. The precision and recall for entity matching are improved further. ⑤Besides the attribute entropy and mutual information approaches used to resolve the difficulties for computing attributes weights, the BP neural network is introduced to entity matching domain in the dissertation. BP neural network can identify the corresponding attributes by analyzing the inner relationship among attributes using the self learning ability of BP network itself, and avoid computing the attribute weight directly. Firstly, the factors to interference the performance and accuracy of neural network are analyzed based on a practical classification problem. Then the entity matching algorithm, modified algorithm, and two-phase-check entity matching algorithm based on BP neural network are proposed respectively. The experimental results show our approaches are very effective. Especially, the two-phase-check algorithm based on BP neural network for entity matching can resolve the problem that different inputs for neural network may have the same outputs effectively and avoid resulting in the error results, and improve the accuracy of entity matching further. Finally, the researches in the dissertation are summarized and the future works are presented.更多还原