一种新的快速特征选择和数据分类方法

针对数据分类问题提出一种新型高效的特征选择和规则提取方法.首先通过减少初始区间数量改进Chi-Merge离散化方法,再采用改进的Chi-Merge离散化连续型特征变量;特征离散化后,统计样本数据在每个特征子集划分下的频数表,并根据频数表计算数据不一致率,再利用顺序前向最优搜索的方法,快速确定特征数量由小到大的每一个最优特征子集;根据特征子集对应的数据不一致率差异最小化原则,完成特征个数最小化的最优特征子集筛选;根据最优特征子集的数据频数表,可直接提取数据分类规则.实验表明,快速提取的规则可获得较好的分类效果.基于该特征选择方法,提出一种面向分布式同构数据的快速分类模型,不但具有良好的分类效果,还支持对样本数据内容的隐私保护.

Novel and Efficient Method on Feature Selection and Data Classification

A novel feature selection method for data classification problems,as well as a quick rule extraction scheme,are proposed in this paper.At first,the Chi-Merge discretization method is improved by reducing the initial intervals.Using the improved method,the continuous attributes can be effectively discretized.After the attributes discretization,all contingency tables on variant feature patterns can be calculated quickly,and the inconsistency rate can also be generated for each contingency table.The key sequential of features can be identified by selecting the minimum inconsistency rate,and the optimized feature subset can also be achieved efficiently based on the sequence forward search strategy.At last,based on the data contingency table under the selected feature subset,the classification rules can be extracted with one-pass.The experiments show that the proposed data classification scheme obtains good performance.Furthermore,the proposed feature selection and rule extraction method can be extended for the classification applications on distributed isomorphic datasets.The proposed distributed classification method is also simple,efficient with high performance,as well as with privacy-preserving property for contents of sample data.