Unsupervised feature selection is critical in high-dimensional data analysis, as it identifies the most informative features without relying on label information. In this paper, we propose a novel robust dual space factorization for unsupervised feature selection (RDSF-UFS) framework based on symmetric nonnegative matrix factorization and nonnegative matrix tri-factorization (NMTF). The proposed method explores sample and feature affinity matrices to capture intrinsic data structures in a dual latent space. Specifically, we introduce a robust objective function incorporating the Frobenius and \(\ell _{2,1}\) -norms to enhance resistance to noise and outliers. We further impose an orthogonality constraint on the latent factor matrices to promote sparsity and improve clustering performance. RDSF-FS uses NMTF to create a bridge or connector that improves the latent representation, thereby strengthening the model’s ability to maintain the structural information within the data. An efficient optimization algorithm is derived to iteratively update the latent factors by solving the associated gradient-based equations. We evaluated the performance of RDSF-UFS on eight benchmark datasets, including biological microarray, face image, speech signal, and digit image datasets. Experimental results demonstrate that RDSF-UFS outperforms state-of-the-art feature selection methods in terms of normalized mutual information and accuracy in specific datasets, highlighting its ability to uncover meaningful feature patterns and improve clustering performance.
