【作者】 张斌；

【导师】 胡学钢； 张晶；

【作者基本信息】 合肥工业大学 ， 计算机应用技术， 2012， 硕士

【摘要】 随着信息技术的快速发展和广泛应用，每天都在产生大量的数据。如何在“数据丰富，但信息贫乏”的环境中发现有价值的、有趣的、具有指导意义知识，是数据挖掘的重要任务。同时，数据挖掘也受到了越来越多研究者的关注。而分类研究作为数据挖掘的重要课题，被广泛地应用于商业决策、医疗诊断等方面。结构化的关系数据作为分类研究的主要对象，通常来源于真实的关系数据。然而，直接从关系数据库中获得的多关系数据，相比较于单一的关系数据而言，更具有天然性。但传统分类方法仅适用于单一关系的数据。因此，为了将传统的分类方法应用于多关系数据中，并做出有效的分类预测，需要建立一个从多关系到单关系的桥梁，将多关系数据有效地转换为传统分类方法能处理的单关系数据。所以，本文基于多关系转化这一主要思想，构造了多关系转化模型，并在此基础上提出了IWT以及MRT两种算法。它们分别利用不同策略，有效地提高了多关系转化的效率，解决了转化过程中的统计偏差及衍生问题，并获得了较好的预测效果。本文的研究工作从以下几个方面展开：(1)建立高效的连接路径。通过分析关系间连接属性的对应关系，利用广度优先遍历方法，重新对连接树进行构造。提高了传递的效率。(2)构建关系选择模型。从多关系全局来看：在多关系数据中，目标关系仅仅有一个，而关系数据库中包含了海量的背景关系。在这些背景关系中，并不是每一个关系都包含了对用户或者分类具有重要意义的属性。因此，为了提高挖掘效率，构建关系选择模型，消除冗余关系是转化过程的重要基础。(3)提出基于多关系数据的特征选择方法。从关系内部看：并不是所有属性都是用户关心的或者对类别有较大区分性的。因此，在多关系的数据中进行特征选择，有利于预测性能和效率提高。(4)转化过程中统计偏差问题的分析和处理。关系间元组的一对多映射和多对一映射，以及转化过程中的空值问题，是引起属性重要性在转化前后不一致的根本原因。因此，本文提出了基于元组转化和实例加权转化两种解决策略，保持转化前后属性的重要性一致。实验证明，转化策略在保持了统计一致性的前提下，利用传统分类方法获得了较好的预测效果。更多还原

【Abstract】 The rapid development of information technology has brought a giganticgrowth of the data. How to mine the worth, interesting and meaning knowledge touser are the important task of Data Mining. Meantime, it is focused on by more andmore researchers. Moreover, classification is the one of the important project ofdata mining. It is applied on the finacial decision making, medical research and soon.Structrue data is the main object of classifications, and it mainly comes fromthe relational data stored in the real world applications. These relational dataobatianed by relational database is more nature than the“flat”data. However, thetraditional classifications are only used on the data represented in single,“flat”relational form. Therefore, a bridge is necessary built bettwen multi-relational dataand single relational data. And then, the traditional classifications can be appliedon the single form data which is transformed by the bridge, and make the effectivepredicting. Based on the multi-relational transformation, we construct themulti-relational transformation model, and design two algorithms IWT and MRT.The algorithms make use of different strategies increasing the efficiency, workingout the statistics bias and the relevant problems and getting the better predictingperformance.The main contributions of this dissertation are as follows:(1) Construct the high efficiency link path. We rebuilt the link path byanalyzing the relation among attributions, which exist on the relations linking eachother. In addition, we make use of the breadth-first search to access the link path.The new link path owns the higher efficience.(2) Construct the selecting model of relations. On the global view, in themulti-relational data, the target relation is only one, but there are huge backgroundrelations in the relational database. However, the attributions, which are meaningfor classficating or the user making decision, are not distributed on everybackground tables. Thus, the selecting model is necessary in the transformation forimproving the mining efficiency and removing the redundancy relations.(3) Construt the feature selection function based on multi-relational data. Onthe local view, all attributes are not focused on by user or important to distinguish the class labels. So, the feature selection based on multi-relational data is useful toimproving the efficiency and predicting performace.(4) Analyze and process the statistics bias in the transformation process. Theone-to-many and many-to-one among the tuples of relations and the null valueexisted in the transformation process is the root cause that leads to be not consistanton evaluating the importance of attributes bettwen before and after conversion ofrelations. Thus, the instance weighting transformation and tuples transformationstrategies are proposed to keep the importance of attributes consistant bettwenbefore and after transformation. Our comprehensive experiments demonstrate thewell predicting performance of our methods, based on keeping the consistence.更多还原