Feature selection has attracted widespread attention with the massive growth of high-dimensional data. In recent years, all kinds of unsupervised feature selection methods have been presented. However, most of these methods can not fully explore the local geometric structure of the original data, which has been proven very important in unsupervised feature selection. To tackle this problem, we present a novel feature selection algorithm called dual-graph subspace learning based feature selection (DGSLFS). Specifically, on one hand, DGSLFS conducts feature selection procedures based on subspace learning, which can guarantee the useful information hidden in the original space be well exploited. On the other hand, we develop two novel graphs on samples and features, respectively, which can well preserve the local geometric structures. In addition, we impose an ${\ell }_{2,1}$-norm to constrain the reconstruction error term and the feature selection matrix. Thus, DGSLFS is robust to outliers and noises, and can guarantee the sparsity of features. The experimental results on several popular datasets show that our proposed algorithm can obtain encouraging results in comparison with some state-of-the-art algorithms.

随着高维数据的大量增长，特征选择引起了广泛关注。近年来，出现了各种无监督的特征选择方法。然而，这些方法大多不能充分探索原始数据的局部几何结构，这在无监督特征选择中已被证明是非常重要的。为了解决这个问题，我们提出了一种新的特征选择算法，称为基于双图子空间学习的特征选择（DGSLFS）。具体来说，一方面，DGSLFS 进行基于子空间学习的特征选择过程，可以保证隐藏在原始空间中的有用信息被很好地利用。另一方面，我们分别开发了两个关于样本和特征的新图，可以很好地保留局部几何结构。此外，${\ell }_{2,1}$-norm 来约束重构误差项和特征选择矩阵。因此，DGSLFS 对异常值和噪声具有鲁棒性，并且可以保证特征的稀疏性。在几个流行数据集上的实验结果表明，与一些最先进的算法相比，我们提出的算法可以获得令人鼓舞的结果。