Latent Space Embedding for Unsupervised Feature Selection via Joint Dictionary Learning
Fan Yang, Jianhua Dai, Qilai Zhang · 2019
With the prevalence of unlabeled data, unsupervised feature selection is vital for comprehensive analysis of unlabeled high-dimensional data. Most existing unsupervised feature selection methods first generate cluster labels by specific techniques and then select features that can preserve cluster structure well. However, the selected features only reflect the distribution information of pseudo label space but ignore that of feature space. Instead, we propose a novel method, latent space embedding for unsupervised feature selection, which considers the common distribution of feature space and pseudo label space, spectral analysis and feature selection simultaneously. Inspired by the success of joint dictionary learning in cross-modality cases, we introduce a latent space shared by feature space and pseudo label space. By utilizing the mapping between feature space and latent space, features that can well maintain the common distribution of features and pseudo labels are selected. The ℓ2,1-norm minimization constraint is added to the objective function to handle outliers and noises. Experimental results on benchmark datasets demonstrate that our algorithm outperforms the compared methods in terms of clustering tasks.