Unsupervised feature selection plays a crucial role in handling high-dimensional data, especially in scenarios where labels are unavailable. This study introduces a novel feature selection method that integrates orthogonally constrained matrix factorization, Hessian regularization, and non-convex sparsity -- l_ (2, 1-2) -norm regularization. The proposed method captures the local geometrical of the data while keeping it easy to interpret and sparse. Using the l_ (2, 1-2) -norm, it highlights important features, removes unnecessary information, and makes the model more resistant to noise and outliers. The use of Hessian regularization preserves the intrinsic manifold structure, and orthogonal constraints promote independence among latent components. Experimental results on multiple datasets demonstrate the effectiveness of the proposed method, achieving competitive performance in clustering accuracy and normalized mutual information (NMI), even with significantly reduced feature dimensions. This study emphasizes the potential of the proposed framework for unsupervised learning tasks involving high-dimensional, unlabeled data.
Dang et al. (Mon,) studied this question.