Multi-Label Classification with Weakly Labeled Data Based on Concept-Cognitive Learning

With the increasing complexity of datasets in reality, it is difficult to obtain instances with full labels, resulting in weak label problems. The existing methods are mainly based on the low rank and instance manifold regularization assumption of label matrix to restore the ground-real label space, while ignoring the influence of semantic noise caused by label missing on the above assumption. Therefore, we propose a method to reconstruct the label space through concept-cognitive learning to solve the problem of multi-label classification with weak label. Specifically, we construct the relationship between the label and the feature by concatenating the label concept and the feature concept, and recover the label space by the similarity between the instance and the concept. Finally, a multi-label classification model is constructed by using label correlation. Experiments on real-world multi-label datasets verify the stability and superiority of the proposed algorithm.