Predicting political sentiments of voters from Twitter in multi-party contexts

Aparup Khatua, Apalak Khatua, Erik Cambria

Abstract: Highlights • Linguistic analysis of social media data for identifying political leanings. • Analysis of mix tweets for predicting political sentiments of social media users. • Employed regression and neural network models for user-classification. Abstract Prior Twitter-based electoral research has mostly ignored multi-party contexts and ‘mix tweets’ that jointly mention more than one party. Hence, we investigate the complex nature of these mix tweets in a multi-party context, and we argue mix tweeting patterns of users implicitly capture their political opinions. We predict the political leaning of users based on their mix tweeting patterns in the context of the 2014 Indian General Election. We have agglomerated 2.4 million tweets from 0.15 million unique users. Next, we employ a multinomial logit regression model to test the hypothesized causal relation between mix tweeting patterns and the political leaning of users. Additionally, we also employ neural network-based algorithms to predict political leaning. Our study demonstrates that user-level mix-tweeting patterns can reveal the political opinions of Twitter users.