Predicting political parties based on the language of their socal media activities

Abstract

In this article we investigated if political parties can be distinguished by the language of their social media activities. We used a simple two-layer neuronal network to predict political parties by their social media posts. The neuronal network was trained on the facebook posts of eight German parties with a simple Naive Discriminative Learning rule.
A cross-validation analysis revealed good accuracy of the predictions for all political parties and additionally showed that the accuracies for the different parties are not homogeneous. A post-hoc analysis of the most activating cues revealed that the best predictors for each party are the names of famous politicians. Furthermore, the post-hoc analysis of the vector semantics of the model showed different patterns of the activation of the cues for two parties.

Structure

This project is structured into three parts. The first part is preprocessing and consists of an R-Script (R Core Team, 2016) which extracts the important informations from the output files of facepager (Keyling & Jünger). The python script preprocess.py preprocesses these preselected informations to an eventfile event_file.tab which must be copied to model/data/.
In the model folder the steps are specified in training.R which mainly runs the cross-validation by training a ndl2 (Arrpe et al., 2015) model on subsets of all events where each training 1000 events were left out. The result of the cross-validation is saved in activations.rda which needs to be copied to analysis/data/.
In analysis the analysis.R script specifies the analysis of the results and plots some results to analysis/plots/.

NOTE: We used this structure over putting all scripts in one folder with one script sourcing them all, because with this structure it is easier to obtain overview over the project.

References

Antti Arppe, Samuel Bitschau, Nathanael Schilling, Peter Hendrix,
    Petar Milin, R. Harald Baayen and Cyrus Shaoul (2015). ndl2: Naive
    Discriminative Learning. R package version 0.1.0.9002.
Keyling, T., & Jünger, J. (2013). Facepager (Version, f.e. 3.3).
    An application for generic data retrieval through APIs. Source
    code available from https://github.com/strohne/Facepager.
R Core Team (2016). R: A language and environment for statistical
    computing. R Foundation for Statistical Computing, Vienna, Austria.
    Retrieved from https://www.R-project.org/.