Improving Social Vote Recommendation in VAAs: The Effects of Political Profile Augmentation and Classification Method

Abstract

Voting Advice Applications (VAAs) are online tools used by voters in order to identify their political stance in relation to parties / candidates running for elections. Traditional approaches are based on some standard vector space distance metrics (e.g. Euclidean Distance), that measure the distance between the political profile of a voter - expressed by her/his answers on a series of policy statements - against those (answers) of parties / candidates. A new paradigm, the so-called Social Vote Recommendation (SVR), extends traditional VAAs with a peer (i.e., voter to voter) opinion matching based on the principles of collaborative filtering. The problem of vote recommendation in that case is equivalent to the problem of matching a multidimensional vector (profile of the current voter) to a set of vectors (profiles of voters that support a particular political party). Previously, this functionality was offered using the Mahalanobis distance; a model that represents the 'average' voter of each party is created, and then, the distance between the active user and the 'average' voter of each party is calculated. In this paper we explore ways in which current best practices can be evaluated and compared to potentially better performing machine learning approaches for use in the domain of VAAs. In addition, we investigate the effects of political profile augmentation with the so-called supplementary questions and we show that users' education level and demographics, such as gender and age, along with the reason of vote choice consistently improve SVR.