Unsupervised Belief Representation Learning in Polarized Networks: A Variational Graph Auto-Encoder Approach

In this paper, we propose an Information-Theoretic Variational Graph Auto-Encoder (InfoVGAE) for polarity representation learning in an unsupervised manner. It jointly learns the belief embedding of both users and their claims in the same latent space. In order to better disentangle the latent space, a total correlation regularizer, a PI controller, and the rectified Gaussian Distribution are adopted to constrain the generated distribution. Experimental results show that the proposed InfoVGAE outperforms the existing unsupervised polarity detection methods, and achieves a highly comparable result with supervised methods in terms of F1 score and purity.