Pmae: Pseudo Multi-Label Attention Ensemble

Multi-label classification has been recognized to be challenging since the complex correlations among labels and the combinatorial nature of the output label space. Motivated by the label dimension reduction and ensemble learning, in this paper, we propose a novel multi-label classification algorithm based on ensemble learning, termed PMAE: Pseudo Multi-Label Attention Ensemble. Different from existing landmark-based methods which reduce the predicted label space by selecting a few real labels as landmarks in a "hard" manner, the proposed PMAE develops the encoder-decoder mechanism to effectively utilize the low-dimensional pseudo-label representation space in a "soft" manner, which can capture the underlying correlations among labels and lead to more robust performance in complex scenarios. Moreover, we conduct the feature embedding based on multiple attention networks integrated, then, exponent and hinge diversity regularization are introduced to ensure the diversity and complementarity of multiple learners. Experimental results on multiple bench-mark datasets show the superiority of our proposed method over other state-of-the-arts.