不确定数据和代价敏感学习研究

发布时间：2018-01-15 10:04

本文关键词：不确定数据和代价敏感学习研究　出处：《西北农林科技大学》2017年博士论文　论文类型：学位论文

【摘要】：传统的数据分类算法需要处理的数据是确定且精确的,然而在现实世界中,由于隐私保护,不精确测量,重复抽样,值缺失等原因,数据普遍存在不确定性。直接使用传统分类算法对不确定数据进行分类效果很差,不能满足应用要求,这使得专门针对不确定数据挖掘技术的研究十分必要。代价敏感学习是一个具有重要意义的研究方向,其充分考虑了各类错误分类的代价差异,通过引入代价概念,使得分类器以最小化代价为目标,有效避免了传统分类器仅关注分类准确率的不足。在很多领域,代价敏感学习模型更符合现实应用需求。本课题以不确定数据和代价敏感学习为研究对象,围绕着如何对不确定数据流进行分类分析,如何提出性能更优的代价敏感算法以及如何对不确定数据进行代价敏感学习等几方面展开研究。主要成果包括:(1)提出了一种针对属性不确定的数据流进行分类学习的集成分类算法EDTU。首先,文中基于不确定决策树算法DTU构建了基分类器。然后,为解决基分类器算法效率不高,不能实时的处理数据流的问题,采用了快速构建决策树的方法来改造基分类器。最后,通过动态集成的方式,构建了一个分类器集合,通过在相似样本上的分类准确率来调整分类器权重,并完成分类器的淘汰与更新。实验结果表明,算法能有效的处理不确定数据流的分类问题,并且在不同参数下表现稳定。(2)提出了一种结合贝叶斯和决策树来进行代价敏感学习的算法CS-NBT。首先,文中定义了误分类代价期望,设计了最小化误分类代价期望的属性选择策略,并根据该策略构建了以最小化误分类代价期望为目的的决策树。然后,在所构建决策树的每个叶子节点上学习贝叶斯分类器,通过结合类别概率与代价矩阵,将选择类别概率最大的问题转换为选择类别代价最小的问题,从而使叶子节点中的贝叶斯分类器代价敏感化。相比起现有基于决策树或贝叶斯的代价敏感学习算法,CS-NBT结合了决策树和贝叶斯的优点,利用了代价敏感决策树的结构能够有效的分裂数据,同时,也通过叶子节点处的贝叶斯分类器利用了那些没有出现在树的路径上的属性所携带的信息。实验结果表明,CS-NBT算法展现了优秀的性能,与CSTree、MetaCost和NBT相比,CS-NBT性能稳定,有效的降低了总代价,并且在不同的参数设定下,算法表现良好。(3)提出了弱化贝叶斯假设的代价敏感学习算法CS_ANDE。首先,结合贝叶斯定理和代价矩阵定义了代价期望,将分类中的最大化可能类别问题转化为最小化期望代价问题。其次,通过假设某些特殊属性不独立,而其它属性相互独立,形成了一个较弱的条件独立假设。然后为减少特定的特殊属性带来的偶然性,将属性集中的每个属性都选取为一次特殊属性并求这些值的均值。最后通过使用MetaCost框架,学习对应的代价敏感分类器。CS_AODE与CS_A2DE的不同在于,在CS_AODE只假设一个属性是特殊属性,而在CS_A2DE中则假设两个属性是特殊属性。实验结果表明CS_AODE和CS_A2DE算法在处理代价敏感学习时展现了优秀的性能,有效的降低了代价,并且在不同参数设定下,算法表现良好。(4)提出了基于决策树的代价敏感不确定分类算法CS-DTU。首先,基于概率势的概念定义了代价增益,通过代价增益来表示属性分裂前后代价的改变。其次,设计了使总代价最小化的属性选择方式,并以该方式建立了代价敏感决策树。然后,采用了不确定决策树中的分类方法来分类新样本。实验结果表明,CS-DTU算法能有效的处理不确定数据的代价敏感分类问题,并且在不确定率变化和代价矩阵变化时表现稳定。(5)提出了基于贝叶斯的代价敏感不确定分类算法CS-UNB。首先,通过结合类别概率与代价矩阵,定义误分类代价期望。然后通过选择不同属性进行测试导致的误分类代价期望的变化,依次选择需要进行测试的属性,其中,不确定信息使用概率势的方法进行处理,最后给出相应算法的详细步骤。然后,对CS-UNB进行扩展,提出了相应的单批测算法SBT-CSUNB。采用贪心算法的思想,找出所有对总代价下降正相关的属性,作为需要进行测试的属性集合。实验结果表明,CS-UNB算法和SBT-CSUNB算法都很强的从不确定数据学习代价敏感分类器的能力,在不确定率和代价矩阵变化时,算法都表现良好,且SBT-CSUNB算法展现了更好的稳定性。
[Abstract]:Traditional classification algorithms of data need to be processed is determined and accurate, but in the real world, because of privacy protection, inaccurate measurement, repeated sampling, missing values and other reasons, the data uncertainty prevails. The direct use of uncertain dataclassification effect of poor traditional classification algorithms can not meet the application requirements, this the study on uncertain data mining technology is very necessary. The cost sensitive learning is an important research direction, which fully considers the error classification of various price differences, by introducing the concept of price, makes the classifier to minimize the cost as the goal, effectively avoids the shortcomings of traditional classifiers only focus on classification accuracy. In many areas, cost sensitive learning model more realistic application requirements. This subject to uncertain data and cost sensitive learning as the research object, around How to determine the classification of data stream analysis, how the proposed cost sensitive algorithm has a better performance and how to uncertain data for cost sensitive learning and other aspects of research. The main results are as follows: (1) an attribute uncertain data stream EDTU. ensemble classification algorithm of classification learning first, not to determine the DTU decision tree algorithm is constructed based on the base classifier. Then, in order to solve the base classifier algorithm efficiency is not high, not real-time data processing flow, using a fast method of constructing decision tree to transform based classifier. Finally, through the way of dynamic integration and construct a classifier by the accuracy rate of classification in similar samples on classifier to adjust the weights, and complete elimination and update the classifier. The experimental results show that the algorithm can effectively deal with uncertain data stream classification problem, And stable performance under different parameters. (2) proposed a cost sensitive learning to combine Bias and the decision tree algorithm CS-NBT. first, this paper defined the expected misclassification cost, design attribute selection strategy to minimize the expected misclassification cost, and according to the strategy is constructed to minimize the expected misclassification cost for the purpose of the decision tree. Then, Bias learning classifier in each leaf node of the decision tree, by combining the categories of probability and cost matrix, selecting the maximum probability of category converted to select categories of minimum cost problem, so that the cost of Bias classifier sensitization leaf. Compared with the existing cost sensitive learning decision tree algorithm based on Bias or CS-NBT, combines the advantages of the decision tree and Bias, using the structure of cost sensitive decision tree can effectively split the number of At the same time, also according to, through the Bias classifier leaf node use information attributes that do not appear in the tree on the path to carry. The experimental results show that the CS-NBT algorithm shows excellent performance, and CSTree, compared to MetaCost and NBT, CS-NBT stable performance, effectively reduce the total cost, and in different settings the parameters, algorithm has good performance. (3) proposed CS_ANDE. cost sensitive learning algorithm to weaken the Bias hypothesis first, combined with the Bias theorem and the cost matrix defines price expectations, will maximize the classification of the possible categories problem into the minimization of the expected cost. Secondly, by assuming that some special properties are not independent, and other attributes are independent of each other, forming a weak conditional independence assumption. Then for the chance to reduce special specific attributes, each attribute set which are selected for a time For these special attributes and values. Finally, by using the MetaCost framework, learning the corresponding cost sensitive classifier.CS_AODE and CS_A2DE is different in CS_AODE only assume that a property is the special attribute, while CS_A2DE assumes that two attributes are special attributes. Experimental results show that the CS_AODE and CS_A2DE algorithm showed good performance in cost sensitive learning, effectively reduce the cost, and in different parameter setting, the algorithm performs well. (4) this paper proposes a decision tree based on cost sensitive uncertain classification algorithm based on probability CS-DTU. first, the concept of potential defines the cost gain, said the price change before and after the split attribute through cost gain to make the design. Secondly, the total cost minimization attribute selection method, and the way to establish the cost sensitive decision tree. Then, the uncertainty in decision tree classification Method to classify new samples. The experimental results show that the CS-DTU algorithm can effectively deal with the uncertainty of cost sensitive classification data, and stable performance in the uncertain rate and cost matrix changes. (5) proposed the cost sensitive Bayesian classification algorithm based on uncertain CS-UNB. first, through a combination of class probability and cost matrix. The definition of expected misclassification cost. Then by choosing different attributes of the misclassification cost test leads to expectations, select properties, the need for testing the method for uncertain information using probability potential for processing, the detailed steps and finally gives the corresponding algorithm. Then, the extension of CS-UNB, put forward the corresponding single batch the measuring method of SBT-CSUNB. using the greedy algorithm to find out all of the total cost, decrease attribute correlation, as the need for testing the attribute set. The experimental results table Ming, CS-UNB algorithm and SBT-CSUNB algorithm are very strong, and learn the ability of cost sensitive classifier from uncertain data. The algorithm performs well when the uncertainty rate and cost matrix change, and the SBT-CSUNB algorithm shows better stability.

【学位授予单位】：西北农林科技大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP311.13

【相似文献】