基于样本重叠与近似马尔可夫毯的特征选择算法

Feature selection based on samples overlapping characteristic and modified Markov blanket

随着信息技术的快速发展,数据中的高维特征极大地增加了产生冗余特征的可能性,冗余特征不仅导致搜索空间增大,而且影响了分类的准确率。针对现有的特征选择算法难以解决高维特征选择问题,提出了基于样本重叠与近似马尔可夫毯的特征选择算法(samples overlapping based modified Markov blanket, SOMMB)。该算法首先融合最大信息系数与改进强近似马尔可夫毯去除冗余特征;其次采用样本重叠策略指导前向搜索的过程,选取相关特征。该算法在10个公开数据集上与目前流行的PGVNS、FCBF-MIC、CFS、mRMR、RF、CBFS、ReliefF以及FFSG算法进行对比实验,SOMMB算法的平均准确率为82.519%,对比FFSG获得的最高准确率提升了4.214%,表明SOMMB算法可以提高分类精度。

With the rapid development of information technology, the high-dimensional features in data greatly increase the possibility of producing redundant features, these redundant features not only cause the expansion of search space, but also affect the classification accuracy. Aiming at the difficulty of existing feature selection algorithms for high-dimensional features, this paper proposed a samples overlapping based modified Markov blanket(SOMMB) algorithm. Firstly, this algorithm combined the maximum information coefficient and the modified strong approximate Markov blanket to remove redundant features. Secondly, it used samples overlapping to guide forward search and select relevant features. This paper compared SOMMB algorithm with PGVNS、FCBF-MIC、CFS、mRMR、RF、CBFS、ReliefF and FFSG algorithms on 10 public datasets. SOMMB can reach an average accuracy of 82.519%, 4.214% higher than the highest accuracy of FFSG. It shows superiority in feature selection regarding to classification accuracy.