Sparse low-redundancy multi-label feature selection with adaptive dynamic dual graph constraints

In recent years, with the introduction of manifold learning, graph-based multi-label feature selection has received much attention and achieved state-of-the-art performance. However, improving the graph quality is still a problem that needs to be solved urgently. In addition, existing methods only focus on correlation learning and ignore the feature redundancy problem. To solve these problems, we design an adaptive dynamic graph learning method (ADG), which obtains high-quality dynamic similarity graphs by constraining the Laplacian rank of similarity graphs. Moreover, high-quality dynamic redundancy graphs can also be obtained using ADG, which can better solve the feature redundancy problem. Then, using the \(L_{2,1}\) $L_{2,1}$ norm as a sparsity constraint, we propose sparse low-redundant multi-label feature selection (SLADG) with adaptive dynamic dual-graph constraints, and design an efficient scheme to optimize SLADG. Finally, experimental results on eleven publicly available data demonstrate that ADG can obtain high-quality dynamic graphs, and relative to existing state-of-the-art methods, SLADG improves overall performance by an average of at least 3.0427% across the seven evaluated metrics.