Adaptive label secondary reconstruction for missing multi-label learning

In multi-label learning, an instance is often associated with multiple labels, posing a challenge in obtaining the complete set of labels. This difficulty arises from the interference of missing information, which existing methods struggle to overcome by reconstructing the original labels only once. Therefore, an adaptive label secondary reconstruction for missing multi-label learning called ALSRML is proposed. First, based on reliable label learning, the observable label information is projected into a soft label matrix. Second, ALSRML reconstructs each soft label with the help of a self-expression model. The two levels of reconstructed labels are able to promote each other, resulting in better recovery of missing labels. Then, k-nearest-neighbor instance correlation is used to guide the soft label matrix in obtaining a reliable structure. Finally, ALSRML utilizes local label correlation and ℓ2,1−2-norm to constrain the feature coefficient matrix to be stable and sparse. ALSRML demonstrates its superiority over seven state-of-the-art comparison algorithms across most missing rates through comparison experiments and statistical tests on fifteen datasets. Notably, it achieves significant performance improvements of about 43%, 50%, 85%, and 20% in the metrics of Ranking loss, One-error, Average precision, and AUC at 90% missing rate. Ablation experiments further validate the effectiveness of label secondary reconstruction in recovering missing labels.