Title: Symbolic data analysis using classification tree
Authors: Asanao Shimokawa - Tokyo University of Science (Japan) [presenting]
Masataka Kuroda - Mitsubishi Tanabe Pharma Corporation (Japan)
Etsuo Miyaoka - Tokyo University of Science (Japan)
Abstract: The focus is on the construction method of classification trees based on multi-valued covariates, which are given as elements of the power set of the covariate space. When the covariates are given as multi-values, it is not possible to apply the classical splitting rules directly for constructing a classification tree. Moreover, it is important how to use the distribution information of individual descriptions in the covariate space of each sample for the construction of the model. To address these problem, we propose a new splitting rule, and examine its performance through simulation studies. In addition, we present an application of this model in reference to the classification problem of molecular data which are represented by atom-pair properties.