Spectral Feature Selection Via Low Rank Decomposition and Local Preservation

Feature selection is an important step in data mining and pattern recognition tasks. In unsupervised cases, feature selection becomes more difficult due to the lack of labels in the samples. Therefore, this paper proposes a spectral feature selection (JLRDLP) algorithm that combines low rank decomposition and local preservation. First, the model predicts pseudo-labels using spectral clustering, which jointly learns the cluster labels and feature selection matrix, enabling the algorithm to select more discriminative features. In order to realize the selection of data features, the feature selection matrix is guaranteed to be sparse by
$\mathbf{L}_{2,1}$
-norm minimization constraint; the projection matrix is decomposed into the product of two low-rank matrices to reveal the global correlation of data features; The idea is to preserve the local geometry of the data when transforming the original data into a low-dimensional space. Finally, the experimental results on multiple real datasets show that the above has better robustness, thus improving the effect of the classification.