基于数据分析的决策支持是企业管理的重要手段。然而，在事务数据及所蕴含的信息日益多元与复杂的环境下，属性间的多重共线问题使基于属性信息熵的信息增益计算结果不够客观、准确，导致建立在属性信息增益基础上的决策树不能客观体现属性的分类作用。为此，针对可量化的属性状态及分类问题，文章针对单目标多属性决策问题，提出了一种基于主分量变换的方法。首先，对训练集的数据按照时间序列计算差分并进行标准化处理，实现无量纲化，使各属性变量的相对变化大小能够进行比较；其次，建立各属性（影响变量）与类别标志（响应变量）的相关矩阵，对所获得的矩阵进行主分量变换，得到与响应变量具有最大相关性的主分量矩阵；最后，通过主分量矩阵与各相关矩阵的相关程度来评价属性的分类效用并依此来构建决策树模型。%Decision support based on data analysis is an important way of enterprise management. However, the transaction data and the information it contained is becoming diverse and complex, the multicollinearity of attributes causes nonobjective and inaccurate information gain of the calculation based on information entropy. Thus, the decision tree based on information gain can not objectively reflects the classification function of attributes .Therefore, focusing on the classification problems and their relevant properties which can be quantified, the paper presents a method based on principal component transformation. The method includes several steps. First, calculating the difference of training set according to the time-varying sequence and normalizing the differences to make the difference dimensionless. Then constructing the correlation matrix between response variables and factors variables and obtaining the principal component matrix by principal component transformation. At last, constructing decision tree model by means of comparing the degree of correlation between the principal component matrix and each correlation matrix .
展开▼
