Steeler nation, 12th man, and boo birds: Classifying Twitter user interests using time series

Abstract

The problem of Twitter user classification using the contents of tweets is studied. We generate time series from tweets by exploiting the latent temporal information and solve the classification problem in time series domain. Our approach is inspired by the fact that Twitter users sometimes exhibit the periodicity pattern when they share their activities or express their opinions. We apply our proposed methods to both binary and multi-class classification of sports and political interests of Twitter users and compare the performance against eight conventional classification methods using textual features. Experimental results using 2.56 million tweets show that our best binary and multi-class approaches improve the classification accuracy over the best baseline binary and multi-class approaches by 15% and 142%, respectively.