一种改进的最小二乘回归子空间分割方法

最小二乘回归(LSR)算法是一种常见的子空间分割方法，由于LSR具有解析解，因此它的聚类性能较高。然而LSR算法是应用谱聚类方法聚类数据，谱聚类方法初始化聚类中心是随机的，会影响后面的聚类效果。针对这一问题，提出一种基于聚类中心局部密度和距离这2个特点的改进的LSR算法（LSR-DC）。在Extended Yale B数据集上进行实验，结果表明，该算法有较高的聚类精度，具有一定的鲁棒性，优于现有LSR等子空间分割方法。

An Improved Subspace Segmentation Method Based on Least Squares Regression

 Least Squares Regression (LSR) is a common approach for subspace segmentation, it is very efficient due to a closed form solution. However, spectral clustering is exploited in LSR to obtain the final segmentation results. The drawback of spectral clustering is that it randomly initializes the cluster centers, which may undermine the subsequent clustering performance. In order to tackle this problem, this paper presents an improved LSR algorithm (LSR-DC) based on two characteristics of cluster centers, i.e. local density and distance. Experimental results on the Extended Yale B database show that LSR-DC is robust and is superior to the existing LSR subspace segmentation methods.