Semi-supervised feature selection with minimal redundancy based on group optimization strategy for multi-label data

With the development of intelligence technology, high-dimensional multi-label data exist in practical applications, which makes multi-label learning a significant challenge. Feature selection can obtain more distinguishable features to enhance recognition ability to address high-dimensional problems. Nowadays, most researchers usually evaluate the relevance between labels and features and the similarity between samples. They only focus on the local characteristics of samples without considering the global characteristics. To solve the above problems, in this paper, a novel feature selection approach for semi-supervised learning with minimal redundancy and group optimization strategy (SFGR) in multi-label scenario is proposed. First, a measure based on the Laplacian score and constrain score is utilized to evaluate the relevance between each feature and label. Meanwhile, the global structure of the data is considered via the creation of graphs and a priori information to obtain the feature subset with the highest relevance. Secondly, an optimization iteration algorithm based on a regularization term combining \(\text {l}_1\) ${\text{l}}_1$ -norm and \(\text {l}_2\) ${\text{l}}_2$ -norm is employed to ensure the sparsity of the feature weight matrix and minimize the redundancy. Moreover, a group optimal strategy is applied as a global search approach to fusion the feature subsets to obtain an approximate globally optimal feature subset. Eventually, experimental results on various multi-labeled datasets show that SFGR can perform better than other algorithms.