一种基于K2结构学习算法的石漠化数据特征选择方法

Rock Desertification Data Feature Selection Method Based on K2 Structure Learning Algorithm

在石漠化信息的分类和提取过程中,冗余特征的存在影响分类器的性能,同时增加计算的复杂度。提出一种基于K2结构学习算法的石漠化数据特征选择方法,该方法通过B IC评分方法得到贝叶斯网络的结构,从中获得类节点的马尔可夫覆盖,继而进行特征选择。同时借用不同评分函数的等价性来确定结构学习时所需的样本数,并且给出了样本数的参考。实验表明,该方法由于结合了样本的分类信息,获得的特征子集是最优的,显著提高了分类精度,降低了计算复杂度。

The redundant features affect the performance of classifier and increase the computing complexity in the classification and extraction of rocky desertification information. A feature selection method is proposed for rock desertification data based on K2 structure learning algorithm, getting Bayesian network structure through Bayesian information criterion（BIC） scoring method, obtaining Markov blanket of class node, and conducting feature selection. It determines the number of samples required for structure learning borrowing the equivalence of different score functions, and gives the number of samples for reference. Experiments show that the feature subset obtained by this method is optimal, and significantly improves the classification accuracy and reduces computational complexity by combining the classification information of samples.