基于矩阵分解填充的无监督特征选择方法

Unsupervised Feature Selection Method based on Matrix Factorization Filling

无监督特征选择(Unsupervised Feature Selection,UFS)是一种应用广泛的大数据降维技术,然而传统的无监督特征选择算法并不适用于不完整数据集。近年来研究不完整数据下无监督特征选择的关键是如何依靠不完整数据中的信息以获得特征的紧凑筛选。针对不完整数据集的信息利用不够完全以及现有方法填充不够准确的特点,提出基于矩阵分解填充的无监督特征选择方法。该方法利用所有已知信息对不完整数据集进行填充,之后利用基于l_(2.1)范数的无监督最大间隔特征选择方法进行特征选择。实验结果表明,该算法提高了聚类精度和填充效果。

Unsupervised feature selection is a widely used dimension-reduction technology for big data. However, traditional unsupervised feature selection algorithms are not suitable for incomplete data sets. In recent years, the key of unsupervised feature selection in incomplete data is how to obtain compact feature screening by relying on information in incomplete data. Aiming at the characteristics of insufficient information utilization of incomplete data sets and insufficient filling of existing methods, an unsupervised feature selection method based on matrix factorization filling is proposed. This method uses all known information to fill the incomplete data set, and then uses the l_(2.1) norm-based unsupervised maximum interval feature selection method for feature selection. Experimental results indicate that the algorithm improves the clustering accuracy and filling effect.