| Summary: | 碩士 === 國立成功大學 === 工業管理科學系碩博士班 === 91 === Model trees are similar to decision trees, while they have a linear regression model at each leaf node for prediction, instead of a majority class for classification. It is a useful method for realistic numeric prediction problems. The growing procedure of the model tree is based on a measure called standard deviation reduction (SDR). The property of the SDR will gather instances with relatively close class values into the same node to derive linear regression models. Growing model trees in this way does not consider the linear relations between attributes and class, hence may distort the meanings of data. Thus, we define a new measure called FAR, which quotes the concept of feature selection and coefficient of determination to consider the linear relations between the attribute values and the class values, to grow model trees. This new scheme hopefully could mine more valuable information for the problems of interest. Our experimental results show that the model trees grown by the FAR achieve almost the same prediction accuracy as the ones grown by the SDR and generally have a smaller size to make the interpretation on the learning results easier.
|
|---|
