Adversarial Partial Multi-Label Learning with Label Disambiguation

Partial multi-label learning (PML), which tackles the problem of prediction models from instances with overcomplete noisy annotations, has recently started gaining attention research community. In this paper, we propose a novel adversarial model, PML-GAN, under generalized encoder-decoder framework for partial learning. The PML-GAN model uses disambiguation network to identify irrelevant labels and map training their disambiguated label vectors, while deploying generative as an inverse mapping vectors data samples in input feature space. overall corresponds minimax game, enhances correspondence features output bi-directional mapping. Extensive experiments are conducted on both synthetic real-world datasets, proposed demonstrates state-of-the-art performance.