Install the app

Install this application on your home screen for quick and easy access when you’re on the go.

Just tap Share then “Add to Home Screen”

2021 Conference of the ECPR Standing Group on Politics and Gender

Learning VAA: Alternative Matching Methods for Voting Advice Applications

Guillermo Romero Moreno
University of Southampton
Enrique Chueca
Kings College London
Guillermo Romero Moreno
University of Southampton
Javier Padilla
City University of New York

One of the main goals of Voting Advice Applications (VAAs) is increasing voters’ political competence by providing users with their preferred political party according to their own preferences. In order to do this, VAAs compare and aggregate the preferences of users and political parties in a set of policy issues. In this process, VAAs’ developers must take some unavoidable decisions which lead to different models of VAAs. While these decisions are taken as to be fair, they usually are not faithful to the reality of the party landscape, as they are not always based on data or solid theories. In this paper, we propose a matching method whose parameters are dynamically adapted with the arriving of user responses by using machine learning techniques. It works in a similar fashion than current methods of matching, although learning the importance and the issue voting theory of each question from the arriving data. These parameters are adapted so as to minimise the discrepancy between the given recommendation and the voting intention stated by some of the users before their filling the questionnaire. Unlike previous machine learning methods, ours places special emphasis on respecting the issue voting theories and produces models that are easily interpretable by experts from the field, facilitating the supervision of the model’s right functioning. We use data from the EU-Vox2014 (hopefully EU-Vox2019) in the 2014 European elections to prove the effectiveness of our model for different European countries, reaching higher performance than previous methods of recommendation as measured by the accuracy, mean rank, and f-score. The learned model is also valuable itself, as it contains information about users issue voting approach and saliency on a per-question basis. We believe that our matching method can be of higher quality than other alternatives due to both its higher performance and interpretability.
Share this page