Multi-label learning of missing labels using label-specific features: an embedded packaging method

Learning label-specific features is an effective strategy for multi-label classification. Existing multi-label classification methods for learning label-specific features face two challenges. One is the incompleteness of the training label data, and the other is that existing techniques generate label-specific features and build multi-label classifiers independently, ignoring the decoupled nature of the two phases. In this paper, we propose an embedding packing method that combines the generation of label-specific features and subsequent model induction to deal with the multi-label classification problem of missing labels. The proposed method first recovers the missing labels using higher-order label correlation while completing the generation of label-specific features in the embedded feature space and finally completes the construction of a multi-label classifier by combining empirical loss minimization and learned label correlation. Moreover, this paper utilizes the kernel expansion technique to ensure that the model can handle linearly inseparable data. We use alternating optimization techniques to solve the potential optimization problem effectively. Comparative analysis experiments are conducted on ten benchmark multi-label datasets to verify that the proposed method is as competitive as other state-of-the-art methods in dealing with missing-label data.