统计独立性的离散化新方法

A NOVEL DISCRETIZATION METHOD FOR STATISTICAL INDEPENDENCE

连续属性离散化方法能够提高数据挖掘和归纳学习等算法的分类预测能力。提出一种统计独立性的离散化新方法,该方法改进了基于卡方统计的区间合并函数,不仅考虑了各对合并区间中卡方自由度对离散化结果的影响,而且考虑了数据类分布的影响,很好地衡量了类-属性之间的相互独立性。实验结果表明,新方法显著地提高了Nave-bayes和SVM分类器的学习精度。

Continuous attributes discretization method can improve the classification and prediction capability of such algorithms as data mining or inductive learning.A novel discretization method for statistical independence is proposed,which improves the chi-square-based interval merging function.It considers not only the influence of chi-square freedom degree in every merged interval pair on discretization results,but also that of data class distribution so that the mutual independence between the class and the attributes are excellently measured.Experimental results show that the new method significantly enhances the classification accuracy of Nave-bayes and SVM classifiers.