中文摘要：

利用局部线性嵌入（LLE）算法中获得局部邻域之间的重构关系与使用最小角回归方法解决L1归一化问题都使用回归方法，针对在通过映射获得低维嵌入空间与通过特征选择获得低维空间上有着一致的思想，提出一种能保持局部重构关系的无监督谱特征选择方法。该方法利用最小二乘法计算样本的邻域重构系数，并用这些系数表示样本之间的关系，通过解决稀疏特征值问题获得能够保持样本间关系的低维嵌入空间，最后通过解决L1归一化问题实现自动特征选择。通过在四个不同数据集上的聚类实验结果证明，该方法能更准确地评价每个特征的重要性，能自动适应不同的数据集，受参数影响更小，可以明显提升聚类效果。

英文摘要：

Both locally linear embedding （LLE） algorithm obtaining reconstruction relations between local neighborhood and minimum angle regression methods solving problems L1 normalization, which both took use of regression method, LLE algo- rithm through mapping to obtain a low dimensional embedding space and minimum angle regression methods through feature se- lection to obtain low dimensional space, they had the same ideas. This paper proposed an unsupervised spectral feature selec- tion method which maintained a partial relationship of the reconstruction. The method first calculated the local reconstruction coefficients by least squares, and used these coefficients to represent the relationships between samples. Then it calculated the low-dimensional embedding space which could keep these relationships between samples by solving a sparse eigen-problem. In the end,it got the automatic feature selection result by solving L1 norm problem. In the results of clustering experiment in four different data sets, this method is more accurate to evaluate the important of each feature, can automatically adapt to different data sets, is less sensitive to parameters, and can improve the clustering effect obviously.