Label correlation for partial label learning

Partial label learning aims to learn a multi-class classifier, where each training example corresponds set of candidate labels among which only one is correct. Most studies in the space have focused on difference between and non-candidate labels. So far, however, there has been little discussion about correlation partial learning. This paper begins with research correlation, followed by establishment unified framework that integrates adaptive graph, semantic maximization criterion. work generates fresh insight into acquisition information from space. Specifically, calculated utilized obtain similarity pair instances After that, labeling confidence for instance updated smoothness assumption two should be similar outputs if they are close feature At last, an effective optimization program solve framework. Extensive experiments artificial real-world data sets indicate superiority our proposed method state-of-art methods.