Sparse low-redundancy multi-label feature selection with constrained laplacian rank

As one of the crucial methods for data dimensionality reduction, multi-label feature selection aims to eliminate irrelevant and redundant features from the data and retain only the most representative subset of features. Although many advanced multi-label feature selection methods have been proposed with state-of-the-art performance, it remains a long-standing challenge for multi-label feature selection to remove irrelevant and redundant features from the data. Besides, the existing graph-based multi-label feature selection methods are limited by the single way of learning similar graphs, leading to a considerable model limitation. To address these issues, we use the \(L_{1}\) $L_1$ norm sparsity constraint to eliminate irrelevant features and introduce the Laplace rank constraint to construct dynamic graphs to improve the sparsity of the model and overcome the model limitation problem. Next, we build a penalty term for eliminating redundant features by combining \(L_{1}\) $L_1$ norm and \(L_{2,1}\) $L_{2,1}$ norm to constrain the learning of the feature weight matrix. Then, we combine it with the linear mapping of instances to ground-true labels. We propose sparse, low-redundancy multi-label feature selection with constrained Laplacian rank (SLCLR). Finally, SLCLR was compared with nine advanced existing methods on thirteen benchmark multi-label datasets, and the experiment results on seven commonly used evaluation metrics all validated the good feature selection performance of SLCLR.