Joint subspace reconstruction and label correlation for multi-label feature selection

High-dimensional multi-label data has become more prevalent in many application domains, presenting difficulties and challenges for multi-label learning. As a result, feature selection has been widely used as an effective dimensionality reduction technique in multi-label learning. However, traditional multi-label feature selection (MLFS) methods map the feature space directly to the corresponding label space based on linear assumptions, which is not reasonable in most cases, and there may be redundant information in the original space that affects the performance of the model. In addition, some multi-label feature selection methods prioritize the correlation between input features and their corresponding labels while ignoring the relationship between labels or consider either local label correlation or global label correlation when both types of label correlation are important for feature selection. To compensate for these shortcomings, we propose a novel MLFS method by joint subspace reconstruction and label correlation for multi-label feature selection (JSRLC). First, we use ridge regression to map the original data space onto a low-dimensional manifold space to preserve the structural information of the data and reduce the negative impact of redundant information, and use l(2,1)-norm for the feature weight matrix to facilitate the feature selection process. Second, reconstruction of the low-dimensional manifold label space using subspace reconstruction terms to learn structural information about the original labels. Finally, manifold learning is used to mine the correlation between labels and ensure the consistency between the global and local structure of the original label space and the low-dimensional manifold space. Moreover, extensive experiments are performed on twenty multi-label datasets, and the results show that the JSRLC outperforms eleven comparative methods.