具有高可理解性的二分决策树生成算法研究

二分离散化是决策树生成中处理连续属性最常用的方法,对于连续属性较多的问题,生成的决策树庞大,知识表示难以理解.针对两类分类问题,提出一种基于属性变换的多区间离散化方法--RCAT,该方法首先将连续属性转化为某类别的概率属性,此概率属性的二分法结果对应于原连续属性的多区间划分,然后对这些区间的边缘进行优化,获得原连续属性的信息熵增益,最后采用悲观剪枝与无损合并剪枝技术对RCAT决策树进行简化.对多个领域的数据集进行实验,结果表明:对比二分离散化,RCAT算法的执行效率高,生成的决策树在保持分类精度的同时,树的规模小,可理解性强.

Constructing Binary Classification Trees with High Intelligibility

Binarization is the most popular discretization method in decision tree generation, while for the domain with many continuous attributes, it always gets a big incomprehensible tree which can't be described as knowledge. In order to get a more intelligible decision tree, this paper presents a new discretization algorithm, RCAT, for continuous attributes in the generation of binary classification tree. It uses simple binarization to solve the multisplitting problem through mapping a continuous attribute into another probability attribute based on statistic information. Two pruning methods are introduced to simplify the constructed tree. Empirical results of several domains show that, for the two-class problem with a preponderance of continuous attributes, RCAT algorithm can generate a much smaller decision tree efficiently with higher intelligibility than binarization while retaining predictive accuracy.