Cluster Structure Preserving Based on Dictionary Pair for Unsupervised Feature Selection
Qilai Zhang, Jianhua Dai · 2018
For unsupervised feature selection(UFS), some rep- resentative techniques are commonly employed to convert un-supervised feature selection to supervised case by generating pseudo labels, such as spectral clustering, matrix factorization and dictionary learning. Most existing studies address these techniques separately. In this paper, we propose an algorithm combining spectral clustering with dictionary learning for un-supervised feature selection. In dictionary learning, we share an intrinsic feature space for features and pseudo labels to ensure the consistency of data distribution. The cluster structure is encoded into dictionary learning to follow the priors of data distribution. Then the projection matrix is accessed from data matrix to the intrinsic feature space, which indicates the discriminative capability of original feature space. The l2,1-norm is imposed on the projection matrix to get ranked features. An alternating minimization algorithm is employed to optimize the proposed model. The experimental results reveal the effectiveness of the proposed method.