Quantifying Political Polarity Based on Bipartite Opinion Networks.

Political inclinations of individuals (liberal vs. conservative) largely shape their opinions on several issues such as abortion, gun control, nuclear power, etc. These opinions are openly exerted inonline forums, news sites, the parliament, and so on. In this paper, we address the problem of quantifying political polarity of individuals and of political issues for classification and ranking. We use signed bipartite networks to represent the opinions of individuals on issues, and formulate the problem as a node classification task. We propose a linear algorithm that exploits network effects to learn both the polarity labels as well as the rankings of people and issues in a completely unsupervised manner. Through extensive experiments we demonstrate that our proposed method provides an effective, fast, and easy-to-implement solution, while outperforming three existing baseline algorithms adapted to signed networks, on real political forum and US Congress datasets.Experiments on a wide variety of synthetic graphs with varying polarity and degree distributions of the nodes further demonstrate the robustness of our approach.