Benedikt Boecking, Carnegie Mellon University
Margeret Hall, Karlsruhe Institute of Technology
Jeff Schneider, Carnegie Mellon University
We aim to predict activities of political nature in Egypt which influence or reflect societal-scale behavior and beliefs by using learning algorithms on Twitter data. We focus on capturing domestic events in Egypt from November 2009 to November 2013. To this extent we study underlying communication patterns by evaluating contentbased and meta-data information in classification tasks without targeting specific keywords or users. Classification is done using Support Vector Machines (SVM) and Support Distribution Machines (SDM). Latent Dirichlet Allocation (LDA) is used to create content-based input patterns for the classifiers while bags of Twitter meta-information are used with the SDM to classify meta-data features. The experiments reveal that user-centric approaches based on metadata can outperform methods employing contentbased input despite the use of well established natural language processing algorithms. The results show that distributions over users-centric meta information provides an important signal when detecting and predicting events.