Non-linear Cost-sensitive Decision Tree for Multi-classification

The motivation of this paper is based on a hypothesis that non-linear decision nodes provide a better classification performance than axis-parallel decision nodes do in many practical problems, such as image classification, and voice classification. The algorithm – MNCS_DT is introduced in this paper to create non-linear splits nodes by novel discriminant analysis in decision tree for multi-classification problem and take costsensitive problem into account when the features are selected. In experiment part, we use four UCI data sets to compare the performance of MNCS_DT and C4.5 CS by costs and error rates. The performance of MNCS_DT is better than C4.5 CS. And eight data sets from UCI are used to compare the performance of three different feature sets measured by accuracy, G-mean, and operation time. The performance of feature set consisting of features that follow multivariate normal distribution and altered information gain values higher than average one is better than two other feature sets in most data sets.