Dual Nonlinear Sparse Feature Selection Method

With the advancement of information technology, the growth in data dimensions poses challenges for data mining and machine learning. Feature selection serves as a critical countermeasure, reducing dimensionality while preserving core information to lay the groundwork for subsequent downstream applications. Existing feature selection methods often focus solely on single-level nonlinear modeling, struggling to capture nonlinear correlations between features and thus limiting their representational capabilities. To address this, this paper proposes DNSFS, a sparse feature selection method that simultaneously accounts for nonlinearity at both the sample and feature levels. This approach utilizes a kernel matrix to capture nonlinear dependencies among features and uncover their inherent association patterns. Concurrently, it constructs a hypergraph structure to model complex relationships between samples, thereby preserving nonlinear features at the sample level. Furthermore, a reasonable regularization strategy balances the weighting of nonlinear information across sample and feature levels, enabling sparse feature selection. Experiments demonstrate that DNSFS achieves superior classification accuracy compared to existing methods. Additionally, ablation studies further validate the effectiveness of the dual nonlinear design.