改进的无监督同时正交基聚类特征选择

提出了一种改进的同时正交基聚类特征选择(Improved Unsupervised Simultaneous Orthogonal Basis Clustering Feature Selection，ISOCFS)方法。为有效地对无标签数据进行特征选择，利用目标矩阵来设计正则化的回归模型。目标矩阵通过正交基聚类，获取投影数据点的潜在聚类中心，引导投影矩阵选择判别性的特征。与先前的无监督特征选择方法不同，ISOCFS并不使用数据点预先计算局部结构信息描述目标函数，而是利用目标矩阵进行正交基聚类直接计算潜在的聚类信息。其次，为了减少噪声信息对估计目标矩阵和投影矩阵的干扰，在先前方法基础上，该方法增加了噪声项。另外，该方法利用简单的优化算法即可求解。最后，通过四个常见的微阵列基因表达数据集及五种最近的无监督特征选择方法进行对比实验，证明了ISOCFS方法可以获得更好的聚类效果。

Improved Unsupervised Simultaneous Orthogonal Basis Clustering Feature Selection

This paper presents an improved Unsupervised Simultaneous Orthogonal Basis Clustering Feature Selection (ISOCFS) method. In order to effectively select features for unlabeled data, a target matrix is used to design a regularized regression model. The target matrix obtains the potential clustering centers of the projected data points by orthogonal base clustering, and guides the projection matrix to select discriminative features. Unlike previous unsupervised feature selection methods, ISOCFS does not use the data points to pre-calculate local structure information to describe the objective function. Instead, it uses the target matrix to perform orthogonal basis clustering to directly calculate the potential clustering information. In addition, in order to reduce the interference of the noise to the estimation target matrix and the projection matrix, the method adds the noise term based on the previous method. Moreover, the method can be solved by using a simple optimization algorithm. Finally, four common microarray gene expression datasets and five recent unsupervised feature selection methods are utilized to demonstrate that the ISOCFS method can obtain better clustering results.