<p>Unsupervised feature selection (UFS) aims to identify the most informative features from high-dimensional data without using label information. Traditional nonnegative matrix factorization (NMF) methods, though effective, typically incorporate only a one-way decoding process and lack the ability to simultaneously capture data structure and discriminative features. In this work, we propose SR–NMF–AE, a discriminative NMF framework with an autoencoder-inspired structure that integrates both encoder and decoder perspectives. We embed K-means clustering into the latent space to enhance the separation between features by capturing cluster structure and incorporate dual graph regularization to preserve manifold structure in both data and feature spaces. Additionally, sparsity and correlation regularizations are applied to ensure interpretability and minimize redundancy among selected features. We derive an efficient optimization scheme based on Karush–Kuhn–Tucker (KKT) conditions and provide convergence analysis. Extensive experiments on 10 benchmark image and microarray datasets demonstrate the superiority of SR–NMF–AE over multiple state-of-the-art UFS techniques. The model also exhibits strong robustness under a wide range of hyperparameter values. SR–NMF–AE presents a practical and effective solution for discriminative, structure-preserving unsupervised feature selection.</p>