4.Multi-view unsupervised feature selection based on graph discrepancy learning

In multi-view learning, unsupervised feature selection plays a vital role in reducing dimensionality while preserving discriminative information distributed across diverse data modalities. Despite notable progress, existing approaches frequently exhibit two key limitations: they often overlook the complementary benefits of integrating global and local structural information, and they inadequately model complex nonlinear relationships or align structural representations across views. To address these challenges, we propose a novel framework, termed Multi-view unsupervised feature selection based on graph discrepancy learning (GDFS). The proposed method jointly constructs global graph structures in a projected low-dimensional space and local graphs in a nonlinear kernel-induced space, effectively capturing both high-level semantic structures and fine-grained neighborhood dependencies. A graph discrepancy term is introduced to explicitly reduce structural discrepancies between global and local representations, thus enhancing consistency and robustness. In addition, a low-rank tensor constraint is applied to the stack of global graphs to uncover high-order correlations across views. A consensus clustering matrix is further learned to provide pseudo-label supervision, which guides the selection of discriminative features. Extensive experiments on six benchmark multi-view datasets demonstrate that GDFS consistently surpasses state-of-the-art methods in terms of clustering performance, thereby confirming its effectiveness, scalability, and generalizability.

Multi-view unsupervised feature selection based on graph discrepancy learning, Y.-W. X, X.-J. X*, X.-L. J, Y.-J. X, Neurocomputing, 2025, 656: 131487.