基于近似Markov Blanket和动态互信息的特征选择算法

针对大量无关和冗余特征的存在可能降低分类器性能的问题,提出了一种基于近似Markov Blanket和动态互信息的特征选择算法。该算法利用互信息作为特征相关性的度量准则,并在未识别的样本上对互信息进行动态估值,利用近似Markov Blanket原理准确地去除冗余特征,从而获得远小于原始特征规模的特征子集。通过仿真试验证明了该算法的有效性。以支持向量机为分类器,在公共数据集UCI上进行了试验,并与DMIFS和ReliefF算法进行了对比。试验结果证明,该算法选取的特征子集与原始特征子集相比,以远小于原始特征规模的特征子集获得了高于或接近于原始特征集合的分类结果。

Feature Selection Algorithm-based Approximate Markov Blanket and Dynamic Mutual Information

To resolve the poor performance of classification owing to the irrelevant and redundancy features, feature selection algorithm based on approximate Markov Blanket and dynamic mutual information was proposed. The algorithm uses mutual information as the evaluation criteria of feature relevance, which is dynamically estimated on the unrecognixed samples. Redundancy features were removed exactly by approximate Markov Blanket. So a small size feature subset can be attained with the proposed algorithm. To attest the validity,we made experiments on UCI data sets with support vector machine as the classifier, compared with DMIFS and RcliefF algorithms. Experiments result suggest that,compared with original feature set,the feature subset size obtained by the proposed algorithm is much less than original feature set and performance on actual classification is better than or as good as that by original feature set.