Unsupervised Belief Representation Learning with Information-Theoretic Variational Graph Auto-Encoders

Abstract

This paper develops a novel unsupervised algorithm for belief representation\nlearning in polarized networks that (i) uncovers the latent dimensions of the\nunderlying belief space and (ii) jointly embeds users and content items (that\nthey interact with) into that space in a manner that facilitates a number of\ndownstream tasks, such as stance detection, stance prediction, and ideology\nmapping. Inspired by total correlation in information theory, we propose the\nInformation-Theoretic Variational Graph Auto-Encoder (InfoVGAE) that learns to\nproject both users and content items (e.g., posts that represent user views)\ninto an appropriate disentangled latent space. To better disentangle latent\nvariables in that space, we develop a total correlation regularization module,\na Proportional-Integral (PI) control module, and adopt rectified Gaussian\ndistribution to ensure the orthogonality. The latent representation of users\nand content can then be used to quantify their ideological leaning and\ndetect/predict their stances on issues. We evaluate the performance of the\nproposed InfoVGAE on three real-world datasets, of which two are collected from\nTwitter and one from U.S. Congress voting records. The evaluation results show\nthat our model outperforms state-of-the-art unsupervised models by reducing\n10.5% user clustering errors and achieving 12.1% higher F1 scores for stance\nseparation of content items. In addition, InfoVGAE produces a comparable result\nwith supervised models. We also discuss its performance on stance prediction\nand user ranking within ideological groups.\n