Abstract

This study establishes an innovative and effective approach for generating new variables and interactions for logistic regression using the two data mining techniques TreeNet and association rules analysis. With TreeNet as the first step in our logistic model building, the new variables are generated by discretizing the quantitative variables. With ASA as the following step, the new interactions are generated from all the original categorical variables and all the newly generated predictors from TreeNet. These newly generated variables and interactions (low- and high-order) are used as candidate predictors to build an optimal logistic regression model. The method is tested on and the results given for four medical datasets—heart disease, heart failure, breast cancer, and hepatitis—with the complete model process presented for the last of these. The results indicate that building a model in this way constitutes a major advance in logistic regression modeling that cannot be achieved using other existing methods.

摘要

本研究建立了一种创新且有效的方法，使用 TreeNet 和关联规则分析这两种数据挖掘技术来生成逻辑回归的新变量和交互作用。以 TreeNet 作为逻辑模型构建的第一步，通过离散化定量变量来生成新变量。使用 ASA 作为后续步骤，新的交互作用是从所有原始分类变量和 TreeNet 中新生成的预测变量生成的。这些新生成的变量和交互作用（低阶和高阶）用作候选预测变量来构建最佳逻辑回归模型。该方法针对四种医学数据集（心脏病、心力衰竭、乳腺癌和肝炎）进行了测试并给出了结果，并为最后一个数据集提供了完整的模型过程。