Disentangled Variational Autoencoder based Multi-Label Classification with Covariance-Aware Multivariate Probit Model

Though these methods can be adapted from single-label predictors, they ignore the correlation Multi-label classification is the challenging task among labels. To improve this, classifier chains [Read of predicting the presence and absence of multiple et al., 2009] stack the binary classifiers into a chain and reuse targets, involving representation learning and the outputs of previous classifiers as extra information to improve label correlation modeling. We propose a novel the prediction of the current label. Followup works extend framework for multi-label classification, Multivariate the classifier chains to recurrent neural networks [Wang Probit Variational AutoEncoder (MPVAE), that et al., 2016] to increase capacity and better model the label effectively learns latent embedding spaces as well correlation. Label ordering is critical to these methods as label correlations. MPVAE learns and aligns two since long-term dependencies are typically weaker than shortterm probabilistic embedding spaces for labels and features dependencies. The model structure also restricts parallel respectively. The decoder of MPVAE takes in computation. Another straightforward method is to find the samples from the embedding spaces and models nearest neighbors in the feature space and assign labels to the joint distribution of output targets under a Multivariate test samples by Bayesian inference [Zhang and Zhou, 2007; Probit model by learning a shared covariance Chiang et al., 2012].