Abstract: Value domain partition methods of continuous attributes are important research in data mining and machine learning. Many discretization methods are proposed, and most tend to discuss the discretization of 1-dimension attribute without considering the relationship among the attributes, which is difficult to get optimal discretization results. This paper proposes a value domain partition method of multiple attributes based on MDL principle. It derives a measurement function of multiple attributes partition by defining model selection of multiple attributes. The paper also designs a reasonable algorithm to find the best discretization result. Performance evaluation and analysis demonstrate that the proposed approach improves the classification and learning ability of Naive Bayes classifier.

Key words: data mining, discretization, Minimum Description Length Principle（MDLP）, Naive Bayes

摘要： 连续属性值域划分方法是数据挖掘和机器学习领域的重要课题。但已有的大量离散化方法倾向于研究一维属性离散化问题，没有考虑多属性之间的相互关系，难于获得最佳的离散化结果。提出一种基于最小描述长度理论的多属性划分方法，通过定义多属性的模型选择问题，推导出多属性划分衡量函数；设计一种合理的算法来寻找最好的离散化结果。性能评价与分析表明，该方法在Naive贝叶斯分类器上有很好的分类学习能力。

关键词: 数据挖掘, 离散化, 最小描述长度理论, Naive贝叶斯