高次判別のためのリスク因子マイニング

抄録

Suppose we have a set of categorical data, some of which might not be exactly categorical but can be regarded as continuously distributed. We want to discriminate each record either to positive or to negative using other observations as explanatory variables. In such a case, it is natural that we want to use two-factor interactions as well as original data. Now the problem arises. How efficiently and thoroughly can we find 'promising' interaction terms? This article presents a statistical point of view to accomplish this task. We make use of multinomial model fitting by AIC. Then the interaction terms will be chosen so that they give a significantly distributions of a response variable from the marginal distribution of the whole data. Then the cross terms will be put into logit/probit type model. An application to medical data is also shown.