Learning label-specific features via neural network for multi-label classification

In multi-label learning, learning specific features for each label is an effective strategy, and most of the existing multi-label classification methods based on label-specific features commonly use the original feature space to learn specific features for each label directly. Due to the problem of dimensionality disaster in the feature space, it may not be the optimal strategy to directly generate the specific feature of the label in the original feature space. Therefore, this paper proposes a multi-label learning framework that joins neural networks and label-specific features. First, the neural network projects the original feature space to a low-dimensional mapping space to learn potential low-dimensional feature space representations, and this nonlinear feature mapping can mine the potential feature information inside the complex feature space. Then, in the low-dimensional mapping space, specific features of the labels are learned using empirical minimization loss. Finally, a unified multi-label classification model is constructed by considering label correlation and instance similarity issues. Extensive experiments are conducted on 12 different multi-label data sets and demonstrate the better generalizability of our proposed approaches.