Label-specific disentanglement and correlation-guided fusion for multi-label classification

In multi-label classification, label-specific features have stronger connection with the label than original ones. Popular methods rely on either representative instance-based transformation or shared feature selection. However, the former hardly achieves good coupling between stages, while the latter lacks pertinence in feature generation. In this paper, we propose an end-to-end approach named label-specific disentanglement and correlation-guided fusion (LSDF) to handle these issues. First, the self-attention mechanism is employed to explore shared feature representations with rich semantic information. Second, the specific features are disentangled from the share one via constructing dedicated multilayer perceptron. Third, guided by third-order label correlation, the tailored features of each label are enriched by fusing with that of the relevant two. Finally, the specific features of each label are separately fed into their respective classifiers for prediction. The results of comprehensive experiments on sixteen benchmark datasets validate the superiority of LSDF against other well-established algorithms. The source code is available at https://github.com/FanSmale/LSDF.