Sparse Discriminant Graph Embedding for Feature Extraction

Linear Discriminant Analysis (LDA) is widely used for dimensionality reduction and enhancing classification. However, traditional LDA methods encounter challenges with large-scale data, such as noise sensitivity and difficulty in determining projection directions. To address these issues, a novel approach named Sparse Discriminative Graph Embedding for Feature Extraction (SDGE) is proposed. Specifically, SDGE incorporates a Laplacian regularization to preserve the local similarity and manifold structure, thereby enabling effective handling of non-linear transformation tasks. Notably, SDGE enhances model controllability and manages algorithm complexity by integrating regularization terms and the constraint conditions into an unified framework. Furthermore, SDGE employs the \({l}_{\text{2,1}}\) -norm to adaptively identify the most discriminative features for discriminant analysis. It also introduces an orthogonal constraint and a sparse constraint simultaneously, ensuring the preservation of key information from the original data while enhancing robustness against noise. Extensive experiments on four real-world databases demonstrate the competitiveness of SDGE against state-of-the-art feature extraction methods.