Unified Local and Global Structure Learning for Feature Selection

Feature selection is an indispensable key technology for tackling the challenges of high-dimensional data. Although prior research indicates that simultaneously preserving local and global data structures provides complementary information, existing methods mostly perform simple linear superposition of the two, making it difficult to characterize their underlying correlations and interactions. This results in structural information being underutilized, thereby limiting the model’s representational capacity and generalization performance. To overcome this limitation, we propose a unified feature selection model that collaboratively learns and aligns local and global structures within a shared low-dimensional subspace. The introduced alignment mechanism dynamically coordinates structural information, ensuring structural consistency during learning and enhancing representation quality. Furthermore, we develop an efficient iterative optimization algorithm and provide a rigorous theoretical proof of its convergence. Extensive experiments on multiple benchmark datasets demonstrate that our method significantly outperforms existing feature selection algorithms in various clustering tasks, which fully validates the model’s effectiveness, robustness, and universality.