Decision tree algorithm in univariate tests causes large-scale,complex rules that are difficult to understand.Multi-variable decision tree is effectively used in the classification of data mining.The key to build it lies in the reasonable choice of attributes combination based on the interconnection between attributes.Based on the rough set theory of attribute dependability and the concept of conditional attributes dispersion degree in information system,a new multi-variable decision tree algorithm called RD is proposed.The results of experiments on the UCI show that the decision tree built by the proposed method has better classification results than those of ID3 algorithm and multi-variate decision tree construction algorithm based on the relative core of attributes.

This paper develops a supervised dimensionality reduction method based on local mean(LMMDA).For every point,LMMDA tries to find the mean of the farthest k1-neighbor points with the same class label and the mean of the nearest k2-neighbor points with the different class label,constructs within-class scatter matrix and between-class scatter matrix,and extracts features through the fisher criterion.This method enlarges the margin of the different class data point.It can retain the true neighborhood structure of the data,and has strong robustness.The experimental results on ORL and Yale face image database show the effectiveness of proposed method.

In this paper, a new decision tree construction algorithm (MIDT) is proposed. MIDT (Multiple Informative Decision Tree) uses principal component analysis to integrate information gain, samples distribution information and correlation coefficient as the basis of the selection of splitting attributes. This method can overcome the disadvantage of ID3 decision tree construction method that uses information gain as the splitting attributes selection criteria as a result of its tendency to select the attribute with more values. And moreover, it can exert the complementarity between decision of entropy mean and decision of samples distribution.The results of experiments on the standard data sets provided by UCI show that the decision tree constructed by MIDT has higher classification accuracy and is more stable than ID3 and parametric estimation decision tree algorithm.

A classifier ensemble method Cagging based on class information was proposed.Training sets of each classifier generated through selecting samples repeatedly based on class information enhanced the diverse of each classifier.The classify results of all classifiers were combined by voting with weight vector to use the diverse of each classifier better.This weight vector was set for each classifier according to its classification performance to each class.The experimental results on face database ORL verify Cagging's validity.Furthermore,the method to generate the classifier in Cagging can generate new ensemble classifier through incremental learning.So,a classifier ensemble method Cagging-I based on increment learning was designed by extending Cagging,and the experimental results verify its validity.