联合过滤式与嵌入式样本的标签库特征选择

Joint filtering and embedded sample label library feature selection

嵌入式样本标签库中存在重复、异常无效特征,导致特征选择正确率较低。为此,提出了联合过滤式与嵌入式样本的标签库特征选择方法。通过计算总体分布概率密度,分析无效特征高斯隶属度。采用基于k近邻算法联合高斯隶属度的高斯正态分布过滤方式,过滤无效特征。设置最大相关最小冗余度特征选择准则,使得标签库特征之间保持最大相关性,过滤错误率持续最小化。定位标签位置,计算标签库两个特征之间水平距离,将空间位置坐标映射到二维平面上,使用粒子群优化算法,依据标签位置选择标签库特征,在粒子群全局最优解的基础上,更新标记位置,从而获得最优解,完成样本标签库特征选择。实验结果可知,该方法特征数据得到有效标记,未出现丢失,最高选择正确率为0.97,具有精准选择效果。

There are duplicate,abnormal,and invalid features in the embedded sample label library,resulting in low accuracy in feature selection.To this end,a tag library feature selection method based on joint filtering and embedded samples was proposed.Analyze the Gaussian membership degree of invalid features by calculating the probability density of the overall distribution.The filtering method of Gaussian normal distribution based on k-nearest neighbor algorithm and Gaussian membership degree is used to filter invalid features.Set maximum correlation and minimum redundancy feature selection criteria to maintain maximum correlation between tag library features and continuously minimize filtering error rates.Locate the tag position,calculate the horizontal distance between the two features of the tag library,map the spatial position coordinates to a two⁃dimensional plane,use the particle swarm optimization to select the tag library features according to the tag position,update the tag position on the basis of the global optimal solution of particle swarm optimization,so as to obtain the optimal solution,and complete the feature selection of the sample tag library.The experimental results show that the feature data of this method has been effectively labeled without any loss,with a maximum selection accuracy of 0.97,and has a precise selection effect.