Semi-supervised multi-label feature learning via label enlarged discriminant analysis

Multi-label learning arises frequently in various domains including computer vision and machine learning and has attracted great interest in the last decades. However, current multi-label classification methods may be deficient in many real applications with following two constraints: (1) lack of sufficient labeled data and (2) high dimensionality in feature space. To address these challenges, in this paper, we propose a new semi-supervised multi-label feature learning algorithm named as label enlarged discriminant analysis. Different from supervised multi-label learning methods, the proposed algorithm can utilize the information from both labeled data and unlabeled data in an effective way. The proposed algorithm enlarges the multi-label information from the labeled data to the unlabeled data through a special designed multi-label label propagation method. Thus, it can take both labeled and unlabeled data into consideration. It then learns a transformation matrix to perform feature learning to reduce the high dimensionality by incorporating the enlarged multi-label information. In this way, the proposed algorithm can preserve more discriminative information by utilizing both labeled and unlabeled data simultaneously. We have analyzed in theory and extensive experimental results are carried out upon several data sets. They all validate the effectiveness of the proposed algorithm.