Balanced Clustering Driven Non-negative Spectral Learning for Unsupervised Feature Selection

In practical data mining applications, dimensionality reduction through unsupervised feature selection is a critical challenge, especially when handling unlabeled, high-dimensional datasets. Existing approaches often rely on spectral analysis for feature selection, yet they fail to adequately address the balance of data distributions and the manifold structures within feature spaces. To overcome these limitations, we propose a novel unsupervised feature selection method, named Balanced Clustering-driven Non-negative Spectral Learning (BCNSL), which integrates manifold learning and balanced clustering into a unified framework. Specifically, BCNSL constructs a feature graph using a sparse transformation matrix, thereby enabling the incorporation of manifold learning into the feature selection process. Additionally, we introduce a balance-aware regularization term to promote the selection of features. To solve the proposed objective function efficiently, we design an optimization strategy based on the ADMM. Experimental results on several benchmark datasets demonstrate the superiority of BCNSL, showcasing significant improvements in clustering quality and distributional balance compared to existing feature selection methods.