Data Mining from Large Data Set by Generalized Tree Regression Model

我们的研究项目涉及树回归模型作为数据挖掘方法。作为有序树回归模型的推广，我们研究了以下问题。(a)协变量作为加性项的推广本文提出了以交互作用效应度为分裂准则，对回归树中每个结点都含有线性回归项的树模型进行估计的方法。该方法使我们能够建立一个简单的树，它比有序树回归模型具有更少的分裂。(b)响应变量分布假设的推广有序树回归模型隐含地假设响应变量服从正态分布。为了将该模型应用于服务时间响应数据，我们提出了假设响应变量服从泊松复合指数分布的方法。(c)树型回归模型的效度研究以日本长期护理保险中的护理需求认证调查的长期护理服务时间数据为例，通过对实证数据的分析，确认了所提出方法的效度。

Our research project dealt with tree regression models as data mining method. We studied the below subjects as generalization of the ordinal tree regression model.(a) Generalization of the covariates use as additive termWe propose the method to estimate tree models which have linear regression terms in each node of regression tree, using degree of interaction effect as splitting criterion. The method enable us to make a simple tree which have fewer splits than ordinal tree regression models.(b) Generalization of distribution assumption of response variableOrdinal tree regression model assume implicitly that response variable distributes as normal distribution. To apply the models to the service time response data, we propose the method under the assumption that response variable has Poisson compound exponential distribution.(c) Varidity study of the tree regression modelWe analyze empirical data to confirm the varidity of proposed methods, for example, Long-term care service time data which was investigated for the Care Needs Certification in Japanese Long-term Care Insurance.The reduction of calculation time remains as future study.