Abstract
Nonnegative matrix factorization (NMF) is an effective dimensionality reduction and representation learning technique that captures the intrinsic structure of nonnegative data by learning low-dimensional, parts-based representations. By incorporating a small amount of label information, semi-supervised NMF can further enhance the discriminative power of the learned representations for clustering tasks. In this paper, we propose an improved graph-regularized discriminative nonnegative matrix factorization model, termed IGDNMF. The proposed model preserves the local manifold structure through graph Laplacian regularization while introducing discriminative constraints that scale and rotate the representations of labeled samples, encouraging them to align along class-specific axes in the learned representation space and thereby improving class separability. Extensive clustering experiments on five real-world datasets demonstrate that IGDNMF consistently outperforms several state-of-the-art semi-supervised NMF methods.
Funding Information
  • Talent Recruitment Research Program of Kunming University (YJL25020)

This publication has 31 references indexed in Scilit: