联合多流形结构和自表示的特征选择方法

Feature Selection Method Combined with Multi-manifold Structures and Self-representation

特征选择是一种通过去除不相关和冗余的特征来降低数据维数和提高后续学习算法效率的数据处理方法。无监督特征选择已经成为维数约简中具有挑战性的问题之一。首先,通过结合特征自表示能力和流形结构,提出了一种联合多流形结构和自表示(Joint Multi-Manifold Structures and Self-Representation,JMMSSR)的无监督特征选择方法。不同于现有的方法,为了更准确地刻画特征的流形结构,引入一种自适应加权策略来融合特征的多个流形结构。然后,提出了一种简单且有效的迭代优化算法来求解JMMSSR方法的目标函数,并利用数值实验验证了优化算法的收敛性。最后,分别在JAFFE,ORL和COIL203个数据集上进行聚类实验,实验结果验证了与现有的无监督特征选择方法相比,JMMSSR方法具有较好的性能。

Feature selection is to reduce the dimension of data by removing irrelevant and redundant features and improve the efficiency of learning algorithm.Unsupervised feature selection has become one of the challenging problems in dimensionality reduction.Firstly,combining self-representation and manifold structure of features,a Joint Multi-Manifold Structures and Self-Representation(JMMSSR)unsupervised feature selection algorithm is proposed.Different from the existing approaches,our approach designs an adaptive weighted strategy to integrate multi-manifold structures to describe the structure of features accurately.Then,a simple and effective iterative updating algorithm is proposed to solve the objective function,and the convergence of the optimization algorithm is also verified by numerical experiments.Finally,experimental results on three datasets(such as JAEEF,ORL and COIL20)show that the proposed approach exhibits better performance than the existing unsupervised feature selection approaches.